System and method for agented play of soccer using multi-modal generative models

Abstract

The disclosed technology involves the use of one or more of soccer match statistics, text insights, predictions (e.g., team and player at the match level, and team at the season level), graphics, video overlays, and player and ball tracking data. Tracking data can be generated using an in-stadium feed or a broadcast feed. The tracking data can be supplemented with event data, which can be provided by an operator or an automated system based on the event related to the given sport within the venue or via a broadcast feed. The tracking data and / or event data can be used to generate insights, such as match statistics, text insights, predictions, graphics, video overlays, etc. Thus, the tracking data and insights generated in accordance with the subject matter disclosed herein can be specific to a given sporting event and / or sport associated with the sporting event.

Description

Cross-references to related applications

[0001] This application claims the benefit of priority to U.S. Provisional Applications Nos. 63 / 574,624 and 63 / 774,286, filed on April 4, 2024 and March 19, 2025, respectively, each of which is incorporated herein by reference. Technical Field

[0002] Various embodiments of the present invention generally relate to machine learning-based techniques for generating sports event data, and more particularly to systems and methods for extracting and processing user input related to sports event data and performing targeted proxy operations based on the user input. Background Technology

[0003] Current generative artificial intelligence (AI) applications focus on tasks that use text to generate images, videos, or audio (or combinations of video and audio). This is achieved by learning mappings from one modality to another using generative AI techniques. This technique also has applications in dialogue, where initial descriptions can be refined to improve the output.

[0004] Unless otherwise stated herein, the materials described in this section are not prior art to the claims of this application, nor are they acknowledged as prior art or suggestions for prior art by virtue of their inclusion in this section. Summary of the Invention

[0005] In some aspects, the technology of this paper relates to a system for performing actions using agent-based artificial intelligence (AI), the system comprising: one or more motion-specific orchestrators; multiple agents, each agent associated with at least one motion-specific orchestrator; a superior orchestrator trained to select one or more motion-specific orchestrators, wherein the motion-specific orchestrators are trained to select one or more agents to perform actions based on one or more user inputs, wherein the superior orchestrator is configured to execute instructions for: receiving one or more user inputs, wherein the one or more user inputs include a description; determining contextual information and intent information associated with the description; selecting one or more motion-specific orchestrators based on the contextual information and intent information associated with the description, wherein each of the one or more motion-specific orchestrators is trained using one or more motion-specific languages, wherein the one or more motion-specific orchestrators are configured to execute instructions for: selecting multiple agents for performing actions associated with the description based on the contextual information and intent information, wherein one or more agents are configured to execute instructions for: retrieving multiple pieces of information based on the contextual information and intent information; determining steps for performing actions associated with the description; and activating agent processes to perform actions based on the determined steps for performing actions.

[0006] In some respects, the technology described herein relates to a system in which one or more user inputs include at least one of text, audio, drawing, or video.

[0007] In some respects, the techniques described herein relate to a system in which determining multiple pieces of contextual and intent information further includes: extracting one or more metadata items related to the description; and determining at least one keyword or tag associated with the description.

[0008] In some respects, the techniques described herein relate to a system in which selecting multiple agents further includes: mapping one or more metadata items to at least one or more event streams; and determining at least one keyword or tag associated with one or more metadata items, which are associated with at least one or more event streams.

[0009] In some respects, the technology described herein relates to a system in which performing actions includes generating one or more motion content based on received information, and wherein one or more agents are configured to execute instructions to match at least one keyword or tag associated with a description with at least one keyword or tag determined to be associated with at least one or more event streams.

[0010] In some respects, the techniques described herein relate to a system in which generating one or more motion content items further includes retrieving one or more content items associated with a subset of at least one or more event streams.

[0011] In some respects, the technology described herein relates to a system in which each of a plurality of agents is trained to retrieve multiple pieces of information relating to one or more motion tracking data and one or more sports event data, and to arrange one or more motion tracking data and one or more sports event data for performing the action.

[0012] In some aspects, the technology described herein relates to a method for performing sports-related actions, the method comprising: receiving one or more user inputs, wherein the one or more user inputs include at least a description; determining multiple contextual and intent information associated with the description; determining one or more sports-specific language models based on the multiple contextual and intent information; determining one or more sports events and tracking data based on the multiple contextual and intent information; retrieving one or more content items associated with the one or more sports events and tracking data using the one or more sports-specific language models; and transmitting the one or more content items for display on a user device.

[0013] In some respects, the techniques described herein relate to a method in which one or more motion-specific languages ​​are trained using one or more general motion attributes and motion-specific attributes.

[0014] In some respects, the technology described herein relates to a method in which general sports attributes include the number of players, field type, and whether it is a team sport or an individual sport.

[0015] In some respects, the technology described herein relates to a method in which sports-specific attributes include starting lineups, ball-handling-based sports, segmented sports, time limits, point distribution, or free-throw-based sports.

[0016] In some respects, the techniques described herein relate to a method in which determining multiple pieces of contextual and intent information further includes: extracting one or more metadata items related to the description; determining at least one keyword or tag associated with the description; and mapping one or more metadata items to at least one keyword or tag.

[0017] In some respects, the techniques described herein relate to a method in which determining one or more motion-specific language models further includes: matching at least one keyword or tag with a general motion attribute and a motion-specific attribute; determining a threshold number of motion-specific attributes identified for each of the one or more motion-specific language models, wherein if the threshold number of motion-specific attributes identified for one of the one or more motion-specific language models is exceeded, then one or more motion-specific languages ​​are selected.

[0018] In some aspects, the technology described herein relates to a non-transitory computer-readable medium configured to store processor-readable instructions, wherein, when executed by a processor, the instructions execute the following operations: a higher-level orchestrator trained to select one or more motion-specific orchestrators, wherein the motion-specific orchestrators are trained to select multiple agents to perform actions based on one or more user inputs, wherein the higher-level orchestrator is configured to execute instructions for: receiving one or more user inputs, wherein the one or more user inputs include a description; determining contextual information and intent information associated with the description; selecting one or more motion-specific orchestrators based on the contextual information and intent information associated with the description, wherein each of the one or more motion-specific orchestrators is trained using one or more motion-specific languages, wherein the one or more motion-specific orchestrators are configured to execute instructions for: selecting multiple agents for performing actions associated with the description based on the contextual information and intent information, wherein one or more agents are configured to execute instructions for: retrieving multiple pieces of information based on the contextual information and intent information; determining steps for performing actions associated with the description; and activating agent processes to perform actions based on the determined steps for performing actions.

[0019] In some respects, the techniques described herein relate to a non-transitory computer-readable medium in which one or more user inputs include at least one of text, audio, drawing, or video.

[0020] In some respects, the techniques described herein relate to a non-transitory computer-readable medium, wherein determining multiple pieces of contextual and intent information further includes: extracting one or more metadata items related to the description; and determining at least one keyword or tag associated with the description.

[0021] In some respects, the techniques described herein relate to a non-transitory computer-readable medium, wherein identifying multiple agents further includes: mapping one or more metadata items to at least one or more event streams; and identifying at least one keyword or tag associated with one or more metadata items, which are associated with at least one or more event streams.

[0022] In some respects, the techniques described herein relate to a non-transitory computer-readable medium in which performing an action includes generating one or more motion content based on received information, and wherein one or more agents are configured to execute instructions to match at least one keyword or tag associated with a description with at least one keyword or tag determined to be associated with at least one or more event streams.

[0023] In some respects, the techniques described herein relate to a non-transitory computer-readable medium in which generating one or more motion content further includes retrieving one or more content items associated with a subset of at least one or more event streams.

[0024] In some respects, the techniques described herein relate to a non-transitory computer-readable medium in which each of a plurality of agents is trained to retrieve multiple pieces of information relating to one or more motion tracking data and one or more sports event data, and to arrange one or more motion tracking data and one or more sports event data for performing actions. Attached Figure Description

[0025] To gain a detailed understanding of the above-described features of the invention, a more specific description of the invention, which has been briefly summarized above, can be obtained with reference to the embodiments, some of which are illustrated in the accompanying drawings. However, it is worth noting that the drawings only show typical embodiments of the invention and should therefore not be considered as limiting its scope, as the invention can allow for other equally effective embodiments.

[0026] Figure 1 A block diagram illustrating a computing environment according to an example embodiment is depicted.

[0027] Figure 2 An exemplary flowchart for selecting agent-based artificial intelligence (AI) according to an example embodiment is depicted.

[0028] Figure 3A and Figure 3B An AI agent according to an example embodiment is described.

[0029] Figure 4 An exemplary flowchart depicts a method for performing actions using agent-based artificial intelligence (AI) according to an example embodiment.

[0030] Figure 5 A flowchart is depicted for a method for generating an interactive display for a motion content data stream according to an example embodiment.

[0031] Figure 6 An example flowchart for generating tracking and / or event data, according to an example embodiment, is depicted.

[0032] Figures 7 to 8 A flowchart is depicted illustrating an exemplary method for generating motion tracking data using a machine learning model according to an example embodiment.

[0033] Figure 9 An example use case for generating motion tracking data in AR / VR and mixed reality applications is described according to an example embodiment.

[0034] Figure 10Another example of generating motion tracking data according to the example embodiment is depicted.

[0035] Figure 11A An example auxiliary query input interface according to one or more embodiments is described.

[0036] Figure 11B Another example of an auxiliary query input interface according to one or more embodiments is depicted.

[0037] Figure 12 An example output interface according to one or more embodiments is described.

[0038] Figure 13 An example graphical output interface according to one or more embodiments is depicted.

[0039] Figure 14 A flowchart for training a machine learning model according to an example embodiment is depicted.

[0040] Figure 15A A block diagram illustrating a computing device according to an example embodiment is depicted.

[0041] Figure 15B A block diagram illustrating a computing device according to an example embodiment is depicted.

[0042] It is worth noting that, for the sake of simplicity and clarity, certain aspects of the accompanying drawings depict general configurations of various embodiments. Descriptions and details of known features and techniques may be omitted to avoid unnecessarily obscuring other features. Elements in the figures are not necessarily drawn to scale; the dimensions of some features may be enlarged relative to other features to improve understanding of the exemplary embodiments. Detailed Implementation

[0043] Various aspects of this invention generally relate to machine learning for sports applications, and more particularly to systems and methods for performing one or more proxy actions using motion-specific data (e.g., tracking data) identified based on user input. Various embodiments can use machine learning models to automatically generate information related to sporting events or perform actions related to sporting events and their subsequent uses. An event can refer to a specific offensive or defensive play, such as a pass or goal, but can also refer to the entire game. As discussed, existing solutions are unable to generate accurate information related to sporting events.

[0044] According to the various aspects disclosed herein, a multimodal sports learning language model (LLM) can receive text, audio, video, or graphics as input from a user. The multimodal sports LLM can use preprocessed event streams to map corresponding metadata to the user input. This information can be used in the multimodal sports LLM to determine generated motion tracking data, which is output to the user or used by one or more motion-specific agents to perform one or more actions. The output information can take the form of visualizations, retrieval systems, analytics, audio and / or text commentary, or combinations thereof. The output information can be event and / or tracking data generated by the multimodal sports LLM, and / or historical event and / or tracking data associated with user input queries. The actions performed can include, for example, generating highlight reels, generating narratives and / or stories, predicting match results, analyzing the referee's influence during a match, simulating player movement throughout a match, comparing the playing styles of one or more players, etc.

[0045] The following non-limiting example is presented for discussion purposes. In this example, the system receives user input for querying a sports event and accesses relevant database records from a database. The database records may include sports-related data associated with the event, such as information about players, teams, and / or leagues. The system determines the intent and contextual information from the query. This information is then mapped to database records to generate or retrieve sports tracking data based on the received query. The system may output information based on the generated sports tracking data or output the sports tracking data to a client device. For example, a user query might involve comparing the playing styles of two or more players in an upcoming match. The system could output a series of text, video, audio, etc., describing each player's playing style (e.g., offensive, defensive, etc.). Furthermore, the system could prepare similar output highlighting the differences between the playing styles of two players.

[0046] The technical advantages of the disclosed technology include improvements to machine learning. For example, some aspects involve determining intent and contextual information from user input, thereby improving the performance, accuracy, and results of the information to be mapped to motion-related data. In doing so, the disclosed technology provides improvements over existing solutions.

[0047] The terminology used above may be interpreted in the broadest and most reasonable manner, even when used in conjunction with a detailed description of certain specific examples of the invention. In fact, the foregoing may even emphasize certain terms; however, any term intended to be interpreted in any limiting manner will be disclosed and specifically defined in the Detailed Description section. The foregoing general and detailed descriptions are exemplary and explanatory only and are not intended to limit the features.

[0048] As used herein, the terms “comprising,” “including,” “having,” “including,” or other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements may include not only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0049] In this invention, relative terms, such as, for example, “approximately,” “substantially,” “roughly,” and “approximately,” are used to indicate possible variations of ±10% in the value.

[0050] The term “exemplary” is used in the sense of “example” rather than “model.” As used herein, unless the context otherwise specifies, the singular forms “a,” “an,” and “the” include plural referents.

[0051] Other embodiments of the invention will be apparent to those skilled in the art upon consideration of this specification and the practice of the disclosed invention. This specification and examples should be considered as exemplary only.

[0052] Figure 1 This is a block diagram illustrating a computing environment 100 according to an example embodiment. The computing environment 100 may include a tracking system 102 (e.g., located at or in communication with one or more components located at the site 106), an organization computing system 104, and one or more client devices 108 communicating via a network 105.

[0053] Network 105 can be of any suitable type, including a separate connection via the Internet, such as a cellular or Wi-Fi network. In some embodiments, network 105 can use a direct connection to connect terminals, services, and mobile devices, such as Radio Frequency Identification (RFID), Near Field Communication (NFC), Bluetooth™, Bluetooth Low Energy™ (BLE), Wi-Fi™, ZigBee™, Ambient Backscatter Communication (ABC) protocol, USB, WAN, or LAN. Because the transmitted information may be personal or confidential, security concerns may require encryption or other protection of one or more of these types of connections. However, in some embodiments, the transmitted information may be less personal, and therefore, the network connection may be chosen for convenience rather than security.

[0054] Network 105 may include any type of computer networking arrangement for exchanging data or information. For example, network 105 may be the Internet, a private data network, a virtual private network using a public network, and / or other suitable connections that enable components in computing environment 100 to send and receive information between components of environment 100.

[0055] Tracking system 102 may be located within field 106 and / or may communicate with components located at field 106 (e.g., electronic communication, wireless communication, wired communication, etc.). For example, field 106 may be configured to host a sporting event including one or more active subjects 112. Tracking system 102 may be configured to capture the movement of one or more active subjects (e.g., players) on the playing field, as well as the movement of one or more other related active subjects (e.g., objects) (e.g., ball, hockey, referee, etc.). In some embodiments, tracking system 102 may be an optical-based system using, for example, multiple fixed cameras, movable cameras, one or more panoramic cameras, etc. For example, a system consisting of six calibrated cameras (e.g., fixed cameras) may be used to project the three-dimensional positions of the players and the ball onto a two-dimensional top view of the playing field. In another example, a combination of fixed and non-fixed cameras may be used to capture the movement of all active subjects on the playing field, as well as the movement of one or more related objects. Using such a tracking system (e.g., tracking system 102), many different camera views of the playing field can be generated (e.g., high sideline view, free throw line view, convergence view, face-off view, end zone view, etc.).

[0056] In some embodiments, the tracking system 102 can be used for broadcast feeds of a given match. For example, the tracking system 102 can be used to generate match files 110 to facilitate broadcast feeds of a given match. In such embodiments, each frame of the broadcast feed can be stored in the match file 110. The broadcast feed can be a feed formatted for broadcast over one or more channels (e.g., broadcast channels, internet-based channels, etc.). The match file 110 can be converted from a first format (e.g., a format output by one or more cameras or a format different from the format output by one or more cameras) and can be converted to a second format (e.g., for broadcast transmission).

[0057] In some embodiments, the game file 110 may also be enhanced with other event information corresponding to the event data, such as, but not limited to, game event information (pass, shot made, turnover, etc.) and contextual information (current score, remaining time, etc.). According to embodiments, the event data may be generated manually or may be generated in real time by a computing system (e.g., within approximately 30 seconds of the event occurring), as discussed herein. The computing system may generate the event data by, for example, analyzing tracking data (e.g., from tracking system 102) and / or one or more other data types, such as video feeds, excitement data, etc. The computing system may utilize machine learning models to determine when a given set of tracking data or a change in tracking data (e.g., a given player movement, object movement, its change, etc.) corresponds to an event (e.g., a score event, a free throw event, a ball possession-based event, a game type event, etc.). Event data may be automatically identified using machine learning, which is trained to receive the game file 110 or a subset thereof as input and output game information and / or contextual information based on the input. According to the techniques disclosed herein, the machine learning model may be trained using supervised, semi-supervised, or unsupervised learning. Machine learning models can be trained by analyzing training data using one or more machine learning algorithms, as disclosed herein. Training data may include match files or simulated match files from historical matches, simulated matches, etc., and may include labeled and / or unlabeled data.

[0058] According to embodiments disclosed herein, event data can be generated based on tracking data and / or content feeds (e.g., in-field video feeds, broadcast feeds, etc.). For example, tracking data can be generated by providing content feeds to one or more machine learning models. One or more machine learning models can identify players and / or objects in the content feed and convert them into digital representations. The digital representations of players and / or objects and their respective positions can be tracked to identify tracking data, such as movement data (e.g., changes in position), changes in movement, trends, etc. This information can be used by a prediction module to make predictions. Tracking data can be analyzed by machine learning models to determine the correlation between tracking data and event types (e.g., goals, passes, match types, etc.). For example, tracking data can be used to determine when the digital representation of an object (e.g., the ball) crosses a scoring object (e.g., the goalpost). Based on this determination, the event type of a goal can be identified. Furthermore, the digital representation of a player who touched the object (e.g., the ball) before the goal-scoring event can be identified as the player who facilitated or otherwise caused the event (e.g., the goal). Therefore, content feeds can be used to generate tracking data, which can be further used to determine event data corresponding to certain sporting events.

[0059] Tracking system 102 can be configured to communicate with organizational computing system 104 via network 105. For example, tracking system 102 can be configured to provide broadcast streams of real-time or near-real-time matches or events to organizational computing system 104 via network 105. As an example, tracking system 102 can provide one or more match files 110 in a first format (e.g., a format corresponding to a component based on tracking system 102). Alternatively or additionally, tracking system 102 or organizational computing system 104 can convert the broadcast stream (e.g., match file 110) from the first format to a second format. The second format can be based on organizational computing system 104. For example, the second format can be a format associated with data storage device 118, as discussed further herein.

[0060] The organization computing system 104 can be configured to process broadcast streams and / or in-field feeds of matches. The organization computing system 104 may include at least a network client application server 114, a tracking data system 116, a data storage device 118, a round-by-round module 120, a filling module 122, and / or a scheduling module 124. Each of the tracking data system 116, the round-by-round module 120, the filling module 122, and the scheduling module 124 may include one or more software modules. One or more software modules may be a set of code or instructions stored on a medium (e.g., the memory of the organization computing system 104) representing a series of machine instructions (e.g., program code) that implement one or more algorithmic steps. Such machine instructions may be the actual computer code that the processor of the organization computing system 104 parses to implement the instructions, or alternatively, may be higher-level instruction encoding that is parsed to obtain the actual computer code. One or more software modules may also include one or more hardware components. One or more aspects of the example algorithm may be executed by the hardware components (e.g., circuitry) themselves, rather than as a result of instructions.

[0061] The tracking data system 116 can be configured to receive broadcast data from the tracking system 102 and generate tracking data based on the broadcast data. In some embodiments, the tracking data system 116 may employ artificial intelligence and / or computer vision systems configured to derive player tracking data from broadcast video feeds.

[0062] To generate tracking data based on broadcast data, the tracking data system 116 may, for example, map pixels corresponding to each player and ball to points, and may transform these points into a semantically meaningful event layer that can be used to describe player attributes. For example, the tracking data system 116 may be configured to record broadcast video received from the tracking system 102. In some embodiments, the tracking data system 116 may also classify each frame of the broadcast video into trackable and untrackable clips. In some embodiments, the tracking data system 116 may also calibrate a moving camera based on trackable and untrackable clips. In some embodiments, the tracking data system 116 may also use skeletal tracking to detect players within each frame. In some embodiments, the tracking data system 116 may also track and re-identify players over time. For example, the tracking data system 116 may re-identify players who are not in the camera's line of sight during a given frame. In some embodiments, the tracking data system 116 may also detect and track objects across multiple frames. In some embodiments, the tracking data system 116 may also utilize optical character recognition (OCR) technology. For example, the tracking data system 116 may utilize OCR technology to extract score information and remaining time information from the digital scoreboard for each frame.

[0063] These technologies help the tracking data system 116 generate tracking data based on broadcast feeds (e.g., broadcast video data). Such tracking data can be a digital representation of actions or events performed during a match and can also include predictive or simulated data, such as automatically filling in gaps in the match footage covered by the broadcast feed. For example, the tracking data system 116 can perform these processes to generate tracking data across thousands of possessions and / or broadcast frames. In addition to these processes, the organization computing system 104 can go beyond generating tracking data based on broadcast video data. Instead, to provide descriptive analysis and useful feature representations for the orchestration module 124, the organization computing system 104 can be configured to map tracking data to a semantic layer (e.g., events).

[0064] The tracking data system 116 can be implemented using a machine learning model. According to the techniques disclosed herein, the machine learning model can be trained using supervised, semi-supervised, or unsupervised learning. The machine learning model can be trained by analyzing training data using one or more machine learning algorithms, as disclosed herein. Training data may include match files or simulated match files from historical matches, simulated matches, historical or simulated feature representations, etc., and may include labeled and / or unlabeled data. Labeled data may include location information, movement information, object information, trends, active entity identifiers, active entity re-identifiers, etc.

[0065] The turn-by-turn module 120 can be configured to receive turn-by-turn data from one or more third-party systems. For example, the turn-by-turn module 120 can receive turn-by-turn feeds corresponding to broadcast video data. In some embodiments, the turn-by-turn data can represent manually generated data based on events that occur during a game. Although the goal of computer vision technology is to capture all data directly from the broadcast video stream, in some cases, the referee is the final decision-maker regarding the outcome of an event. For example, in basketball, whether a shot is a two-pointer or a three-pointer (or valid, traveling, defensive / offensive foul, etc.) is determined by the referee. Therefore, to capture these data points, the turn-by-turn module 120 can utilize machine learning outputs and / or manually annotated data that reflects the referee's final decision. Such data can be referred to as a turn-by-turn feed.

[0066] To aid in identifying events in the generated tracking data, the tracking data system 116 can merge or align the round-by-round data with the newly generated tracking data (which may include game and time fields). The tracking data system 116 can utilize a fuzzy matching algorithm that combines round-by-round data, optical character recognition data (e.g., shot clock, score, remaining time, etc.), and game / ball position (e.g., raw tracking data) to generate aligned tracking data.

[0067] Once aligned, the tracking data system 116 can be configured to perform various operations on the aligned tracking system. For example, the tracking data system 116 can use per-possession data to refine the positions of the player and ball, as well as the precise frames of end-of-possession events (e.g., shot / rebound positions). In some embodiments, the tracking data system 116 can also be configured to automatically detect events from the tracking data. In some embodiments, the tracking data system 116 can also be configured to enhance events with contextual information.

[0068] For automated event detection, the tracking data system 116 may include a neural network system trained to detect / refine various events in a sequential manner. For example, the tracking data system 116 may include an actor-action attention neural network system to detect / refine one or more of the following: shooting, scoring, points, rebounding, passing, dribbling, free throws, fouls, and / or ball possession. The tracking data system 116 may also include a large number of specialized event detectors trained to identify higher-level events. Exemplary higher-level events may include, but are not limited to: tactics, transitions, pressure, crosses, getting past defenders, back-to-the-basket moves, drives, isolation plays, on-ball screens, offsides, handoffs, off-ball screens, etc. In some embodiments, each of the specialized event detectors may represent a neural network specifically trained to identify a particular event type. More generally, such event detectors can utilize any type of detection method. For example, specialized event detectors may use neural network methods or another machine learning classifier (e.g., random decision forest, SVM, logistic regression, etc.).

[0069] While mapping tracking data to events can capture player representation, the tracking data system 116 can generate contextual information to enhance detected events in order to further construct the best possible player representation. Exemplary contextual information may include defensive matchup information (e.g., who is guarding whom in each frame, defensive formation), as well as other defensive information such as defensive coverage of screens or pressures on the ball.

[0070] In some embodiments, to measure influence, the tracking data system 116 may use a metric called an "influence score." An influence score captures a player's influence on every other player on the opposing team and ranges from 0 to 100. In some embodiments, the value of the influence score may be based on motion principles such as, but not limited to, proximity to a player, distance to a scoring target (e.g., the basket, goal, boundary, etc.), gap closure rate, passing routes, routes to the scoring target, etc.

[0071] The imputation module 122 can be configured to use mean regression to create new player representations to reduce random noise in the features. For example, one of the deep challenges of modeling using only data from a potentially limited number of matches per player (e.g., 20-30 matches) may be the high variance of low-frequency events seen in the tracking data. Therefore, the imputation module 122 can be configured to utilize an imputation method, which could be a weighted average between the observations and the sample mean.

[0072] Therefore, for each player, the tracking data system 116, the round-by-round module 120, and the filling module 122 can work together to generate a raw dataset and a filled dataset for each player.

[0073] The orchestration module 124 can be configured or trained to generate connections and / or associations within cues and user input (e.g., audio, speech, graphics, video, etc.) in multimodal sports LLM to determine appropriate individual orchestrators (e.g., sport-specific orchestrators) and agents to perform one or more actions. The orchestration module 124 may include a parent orchestrator, multiple sport-specific orchestrators, and multiple agents associated with one or more of the sport-specific orchestrators. According to one embodiment, each agent may be associated with a single sport-specific orchestrator, such that each agent associated with a given sport-specific orchestrator is trained based on training data associated with the corresponding sport. The orchestration module 124 can be configured to receive user input (e.g., audio / speech) requesting information related to a specific match (e.g., Manchester United vs. Liverpool) or sport (e.g., football, basketball, American football, rugby, tennis, golf, individual sport, team sport, etc.). The orchestration module 124 can generate one or more connections and / or associations (e.g., contextual information and intent information) based on the user input and event stream (e.g., the Manchester United vs. Liverpool match). Based on the generated connections, the orchestration module 124 can be configured to determine event data associated with the event stream and user input.

[0074] The orchestration module 124 can be configured to select one or more motion-specific orchestrators based on one or more connections and / or associations. Each motion-specific orchestrator can be trained using one or more motion-specific languages. The one or more motion-specific orchestrators can be configured to select multiple agents based on connections and / or associations to perform one or more actions associated with user input. The multiple agents can be configured to retrieve multiple pieces of information based on connections and / or associations and determine the steps for performing one or more actions associated with user input. After receiving information from the multiple agents, the orchestration module 124 can output (e.g., one or more graphics, text, audio, or combinations thereof) the results of one or more actions performed by the multiple agents based on the determined connections and / or associations with user input.

[0075] In some embodiments, the parent orchestrator of orchestration module 124 may include a separate mapping model tailored for each input type (e.g., audio, text, drawing, video, etc.). Given that each input is very different from the others, a single mapping model may sometimes struggle to determine connections and / or associations. In such scenarios, one or more separate mapping models may be employed for a single user input. For example, upon receiving user input (e.g., speech and drawing), orchestration module 124 may utilize one or more mapping models for each received input type. One or more mapping models may determine one or more connections and / or associations based on the received input. Based on the determined one or more connections, orchestration module 124 may output one or more graphics and text and / or perform actions corresponding to the user input (e.g., by activating a corresponding motion-specific orchestrator and a corresponding agent). Referring to the diagrams discussed below (e.g., Figures 2 to 8 Further discussion on orchestration module 124.

[0076] Data storage device 118 can be configured to store one or more competition files 126. Each competition file 126 may include video data for a given competition. For example, the video data may correspond to multiple video frames captured by tracking system 102, tracking data derived from broadcast video generated by tracking data system 116, round-by-round data, augmented data, and / or padded training data. Competition file 126 may be based, for example, on competition file 110 as discussed herein. Competition file 126 may have a different format than competition file 110. For example, a first format of competition file 110 or a subset thereof may be transformed into a second format of competition file 126. The transformation may be performed automatically based on the type and / or content of the first format and the type and / or content of the second format.

[0077] Client device 108 can communicate with organizational computing system 104 via network 105. Client device 108 can be operated by a user. For example, client device 108 can be a mobile device, tablet, desktop computer, or any computing system with the capabilities described herein. Users can include, but are not limited to, individuals, such as subscribers, clients, potential clients, or customers of entities associated with organizational computing system 104, such as individuals who have received, will receive, or may receive products, services, or advice from entities associated with organizational computing system 104.

[0078] Client device 108 may include at least one application 130. Application 130 may represent a web browser that allows access to websites, or it may be a standalone application. Client device 108 may access application 130 to access one or more functions of organizational computing system 104. Client device 108 may communicate via network 105 to request web pages, for example, from network client application server 114 of organizational computing system 104. For example, client device 108 may be configured to execute application 130 to access one or more defined connections and / or associations generated by orchestration module 124 based on user input. Content displayed to client device 108 may be transferred from network client application server 114 to client device 108 and subsequently processed by application 130 for display via graphical user interface (GUI) of client device 108.

[0079] Prediction Engine

[0080] The prediction engine can be configured to predict a team's basic formation. Mathematically, the goal of the role alignment process might be to find transformations... It can An unstructured set of player trajectories Mapped to individual character trajectories An ordered set (e.g., a vector). For the active subject and frame In other words, each player's trajectory can be an ordered set of positions. In some embodiments, This can represent the optimal permutation matrix that achieves this sorting. The goal of the prediction engine might be to find the most probable set of two-dimensional (2D) probability density functions. :

[0081] In some embodiments, the equation can be transformed into one of the entropy minimizations, where the objective is to reduce (e.g., minimize) the overlap (e.g., KL divergence) between the roles. Therefore, in some embodiments, the final optimization equation expressed in terms of total entropy H may become:

[0082] The prediction engine can include a formation discovery module, a role assignment module, and a template module, each corresponding to a different stage of the prediction process. The formation discovery module can be configured to learn a distribution that maximizes the likelihood of the data. The role assignment module can be configured to map each player position to a "role" distribution in each frame. Once the data is aligned, the template module can be configured to map each learned formation to a formation clustering template.

[0083] The organizational computing system can receive tracking data and / or event data for multiple events across multiple seasons or across a single game. For each event, a preprocessing agent can segment the event into multiple segments based on event information. In some embodiments, the preprocessing agent can segment the event into multiple segments based on a variety of events that may occur throughout the game. For example, the preprocessing agent can segment the event into multiple segments based on one or more events, including but not limited to red cards, ejections, technical fouls, flagrant fouls, player disqualifications, substitutions, halftime, innings, quarters, overtime, etc. Typically, each of the multiple segments associated with an event can include intervals of necessary duration (e.g., at least one minute of play, at least two minutes of play, etc.). Such necessary durations allow the organizational computing system to detect team formations.

[0084] Each segment can include a set of tracking data associated with it. Player tracking data can be captured by a tracking system, which can be configured to record the player's movements at a high frame rate (e.g., 10Hz). Location. In some embodiments, player tracking data may also include single-frame event labels (e.g., pass, shot, cross) in each frame of the player tracking data. These frames may be referred to as “event frames”. As shown in the figure, the initial player tracking data can be represented as A collection of player trajectories For the main body of the activity and frame In other words, each player's trajectory can be an ordered set of positions. .

[0085] In some embodiments, a preprocessing agent can normalize the raw player position data. For example, a preprocessing agent can normalize the raw player position data in each segment so that all teams in the player tracking data attack from left to right and have a mean of zero in each frame. This normalization may result in the removal of translation effects from the data. This can produce ensemble results. .

[0086] In some embodiments, the preprocessing agent can initialize the cluster centers of the normalized dataset used for formation discovery with average player positions. For example, average player positions can be derived from a set of... ={ The preprocessing agent can take the average position of each player in the normalized data and initialize the normalized data based on the average player position. This initialization of the normalized data based on the average player position can be used as the initial role for each player to minimize the data variance.

[0087] The organizational computing system can learn a pattern template from the tracking data for each segment. For example, the pattern discovery module can learn a distribution that maximizes the likelihood of the data. The pattern discovery module can structure the initial data into a single... vector, where It can represent the total number of frames. It can represent the total number of players involved in an activity (e.g., ten outfield players in football, five players in basketball, fifteen players in rugby, etc.) and It can represent the dimensions of data (e.g., ).

[0088] The formation discovery module can then initiate a formation discovery algorithm. For example, the formation discovery module can initialize the K-means algorithm using the average position of the players and execute it until convergence. Executing the K-means algorithm until convergence produces better results than the conventional method of running a fixed number of iterations.

[0089] The formation discovery module can then initialize a Gaussian Mixture Model (GMM) using the cluster centers from the last iteration of the K-means algorithm. By parameterizing the distribution to K Gaussian mixtures (where K equals the number of "roles," typically also N, i.e., the number of players), the formation discovery module is able to identify patterns in the data. The optimal formation for maximizing likelihood. In other words, the GMM can be configured to recognize... ,in It can represent making data The optimal formation is the one that maximizes the likelihood. Therefore, the formation discovery module can use GMM clustering instead of stopping the process after the last iteration of the K-means algorithm, because elliptic clustering captures the shape of each player role better than K-means clustering, which only captures the spherical nature of the data cloud for each role.

[0090] Furthermore, it is known that GMMs suffer from component collapse and can get trapped in ill-conditioned solutions. This collapse can lead to irrational clustering, for example, meaningless outputs that may be unusable. To address this issue, the pattern discovery module can be configured to monitor the eigenvalues ​​of each of the components or parameters of the GMM throughout the expectation maximization process. If the pattern discovery module determines that the eigenvalue ratio of any component becomes too large or too small, the next iteration can run a soft K-means update (e.g., a Gaussian mixture with spherical covariance) instead of a full covariance update. Such a process can be performed to ensure that the final clustering output is reasonable. For example, the pattern discovery module can monitor how the parameters of the GMM converge; if the parameters of the GMM are unstable (e.g., “out of control”), the pattern discovery module can identify this unstable behavior and subsequently use a soft K-means update to slowly restore the parameters to within the solution space.

[0091] Hash table / tactical manual learning

[0092] For retrieval tasks involving large amounts of data, one embodiment of the system requires the use of hash tables. Hash tables group similar tactics together so that when a query is made, only the "most likely" candidates are retrieved. Local comparisons can then be made between the candidates, and each tactic within these groups is sorted by the degree of similarity. Previous systems attempted to cluster tactics into similar groups using only a single attribute, such as the trajectory of the ball. However, by using additional information, such as information about players (e.g., identity, trajectory, etc.) and events (passing, dribbling, shooting, etc.) as well as contextual information (e.g., whether the team is winning or losing, remaining time, etc.), the semantics of the tactics can be captured more accurately. Therefore, an embodiment of this system utilizes information about the trajectory of the ball and players, along with game events and context, to create hash tables, thereby effectively learning a "tactical manual" representing the tactics of a team or player. The tactical manual is learned by selecting classification metrics that indicate interesting or distinctive tactics. Suitable categorical metrics could include predicting the probability of scoring in football or basketball (e.g., expected score (“EPV”) or expected goal value (“EGV”). For performance variables, other predictors could also be chosen, such as pass probability, shot probability, probability of moving in a certain direction / trajectory, or player fatigue / injury probability.

[0093] Categorical metrics are used to learn decision trees, which are a coarse-to-fine hierarchical approach where, at each node, a question is posed that groups the data. The advantage of this approach is that it is interpretable and multi-layered, and can act as a "latent factor."

[0094] Bottom-up approach

[0095] In one embodiment of the system, a bottom-up approach is used to learn the decision tree. Various features are used sequentially to differentiate tactics (e.g., first the ball, then the player closest to the ball, then the defender, etc.). By aligning trajectories, reference points can be provided for the trajectories relative to their current positions. This allows for asking more specific questions while maintaining generality (e.g., if a player is in the "point guard" role, what are his / her distances to his / her teammates in the "shooting guard" role, and his / her distances to the defenders in the "point guard" role). Using this approach avoids the need for an exhaustive examination of all distances, the number of which is enormous for both basketball and soccer.

[0096] Top-down approach

[0097] In another embodiment of the system, a top-down approach is used to learn the decision tree. In the first step, all tactics are aligned to a template set. Starting from this initial template set, tactics are assigned to a set of K groups (clusters) using all ball and player information, thus forming the first layer of the decision tree. Backpropagation is then used to prune unimportant players and divide each cluster into sub-clusters (the second layer). This process continues until the leaves of the tree represent a dictionary of tactics that are predictive for a specific task, such as scoring a goal (the third layer).

[0098] Personalization using latent factor models

[0099] In addition to the raw trajectory information, in this embodiment of the system, tactics in the database are also associated with game event information and contextual information. The game events and context of the tactics in the database can be inferred directly from the raw position tracking data (e.g., shot made or missed) or can be manually entered. Player role information can also be inferred from position tracking data or entered separately. In this embodiment of the system, the database model can then be trained by constructing features encoding game-specific information based on position and game data, and then calculating predicted values ​​(between 0 and 1) relative to a classification metric (e.g., expected score).

[0100] Given a sufficient number of examples, the database model can use those examples to personalize for specific player or match situations. However, in practice, specific player or match situations may not be adequately represented by tactics in the database. Therefore, embodiments of the system find examples similar to the situation of interest—whether it's finding players with similar characteristics or teams that play in a similar manner. Using a more general representation of players and / or teams, each player or team is thus represented as a distribution of specific attributes, rather than using explicit team identities. Embodiments of the system use tactics from a hash table / tactical manual learned through the distributed clustering process described above.

[0101] Furthermore, while various aspects have been discussed in relation to individual sports, these aspects described are merely illustrative examples. The disclosed techniques are by no means limited to any particular sport. For example, the present aspects can be implemented for use in other sports or activities such as football, American football, basketball, baseball, hockey, cricket, rugby, tennis, etc.

[0102] Machine learning for team / player prediction

[0103] According to embodiments disclosed herein, a transformer neural network can receive inputs (e.g., tensor layers), where each input corresponds to a given player, team, or match. The transformer neural network can output predictions for one or more given players or teams based on such inputs. More specifically, the transformer neural network can output such generated predictions for a given player or team based on inputs associated with that player or team and further based on the influence of one or more other players or teams. Thus, as discussed herein, predictions provided by a transformer neural network can take into account the influence of multiple players and / or teams when outputting predictions for a given player and / or team.

[0104] The systems described herein may include machine learning systems configured to generate one or more predictions. In some examples, the system may incorporate transformer neural networks, graph neural networks, recurrent neural networks, convolutional neural networks, and / or feedforward neural networks. The system may implement a series of neural network instances (e.g., feedforward network (FFN) models) connected via a transformer neural network (e.g., a graph neural network (GNN) model). Although transformer neural networks are generally discussed herein, it should be understood that any applicable GNN or other neural network that can be graphically interpreted can be used to perform the techniques discussed herein with reference to transformer neural networks.

[0105] Transformer-based neural networks can include a set of linear embedding layers; a transformer encoder; and a set of fully connected layers. The linear embedding layers map the component tensors of the received input to tensors with a common feature dimension. The transformer encoder performs attention along the time and active agent dimensions. The set of fully connected layers embeds the output from the final transformer layer of the transformer encoder to a tensor with the requested feature dimension for each target metric.

[0106] Transformer-based neural networks can be configured to receive input features through this set of linear embedding layers. The input features can be received at different resolutions and time series. The input features can be related to player features, team features, and / or match features. The input features can be fed as tuples of input tensors into the linear embedding layers. For example, a tuple of three tensors can be provided, where the first tensor corresponds to all players in the match, the second tensor corresponds to the two teams in the match, and the third tensor corresponds to the match state.

[0107] Examining this set of linear embedding layers, each layer can contain linear blocks for each input tensor of a tuple, and each block maps the input tensor to a tensor with a common feature dimension D. The output of the linear embedding layer can be a tuple of tensors with a common feature dimension, which can be concatenated along the time and activity subject dimensions to form a single tensor.

[0108] A transformer encoder can be configured to receive a single tensor derived from a linear embedding layer. The transformer encoder can be configured to learn embeddings configured to generate predictions for multiple actions of each agent (e.g., each player and / or team). The transformer encoder can include a series of axial transformer encoder layers, each layer alternately applying attention along the temporal and agent-related dimensions. The transformer encoder can include layers that alternate between applying attention temporally to a sequence of action events and applying attention spatially across the group of players and teams at each event time step. The transformer encoder can include axial encoder layers configured to receive tensors derived from linear layers and apply attention along the temporal dimension and subsequently along the agent-related dimension.

[0109] The attention mechanism implemented by the transformer encoder layer can be graphically interpreted on a dense graph, where each element is a node, and the attention mask is the inverse of the adjacency matrix of the edges between nodes (therefore, the absence of an attention mask implies a fully connected graph). In the case of axial attention used here, with the attention mask in the time (row) dimension, the nodes in the graph can be arranged in a grid, and each node can be connected to all nodes in the same column, as well as to all previous nodes in the same row. In this case, attention may be message passing, where each node can receive messages depicting the state of the nodes near it and subsequently update its own state based on these messages. This attention scheme might mean that when making predictions for a particular player, the model might consider (i.e., pay attention to): nodes containing the player's previous states along the time series; and state nodes of other players, teams, and the current match state at the current time step. Nodes may not need to be homogeneous except for having the same feature dimensions, and therefore the node representing a player can receive messages from nodes representing their team or from nodes representing the player's strength. Thus, the model can learn the interactions between actors and ensure consistent predictions for each actor along the time series. The output of the transformer encoder layer can be a tensor (e.g., an output embedding).

[0110] The final layer of a transformer-based neural network can be a fully connected layer. These layers can embed the output of the final transformer layer of the transformer encoder and map it to the feature dimension of each target metric. The final layer can output a target tuple containing tensors for each of a set of modeled actions for each player and / or team. For example, modeled actions can be empirical distribution estimates of motion statistics, such as shots, goals, passes, etc.

[0111] Training a transformer-based neural network can involve selecting an appropriate loss function for the distributional assumption of each output target. For example, the loss function might be the negative log-likelihood of a Poisson distribution, the binary cross-entropy of a Bernoulli distribution, etc. During training, the loss can be calculated based on the ground truth for each target in the training set, and the loss values ​​can be summed. The optimizer can then be used to update the model's weights based on the total loss. The learning rate can be adjusted according to a schedule with cosine annealing (no warm restart).

[0112] As used herein, a "machine learning model" typically comprises instructions, data, and / or a model configured to receive input and apply one or more of weights, biases, classifications, or analyses to the input to generate an output. For example, the output may include a classification of the input, an input-based analysis, a design, process, prediction, or recommendation associated with the input, or any other suitable type of output. Machine learning models are typically trained using training data, such as empirical data and / or samples of input data, which are fed into the model to build, tune, or modify one or more aspects of the model, such as weights, biases, criteria used to form classifications or clusters, etc. The various aspects of a machine learning model can operate on the input linearly and in parallel via a network (e.g., a neural network) or via any suitable configuration.

[0113] The execution of a machine learning model may include deploying one or more machine learning techniques, such as generative learning, linear regression, logistic regression, random forests, gradient boosting machines (GBM), deep learning, graph neural networks (GNNs), and / or deep neural networks. Supervised and / or unsupervised training can be employed. For example, supervised learning may involve providing training data and labels corresponding to the training data, e.g., as ground truth values. Unsupervised methods may include clustering, classification, etc. K-means clustering or K-nearest neighbors, which can be supervised or unsupervised, may also be used. A combination of K-nearest neighbors and unsupervised clustering techniques may also be used. Any suitable type of training can be used, such as stochastic, gradient boosting, random seeding, cyclic, round-robin, or batch-based training.

[0114] While several examples in this document relate to certain types of machine learning, it should be understood that the techniques according to the invention are adaptable to any suitable type of machine learning. It should also be understood that the examples above are merely illustrative. The techniques and methods of the invention may be suitable for any suitable activity.

[0115] Sports machine learning models

[0116] As discussed in this paper, one or more machine learning models can be trained to understand sports language. Therefore, the machine learning models disclosed herein are sports machine learning models. Such sports machine learning models can be trained using motion-related data (e.g., tracking data, event data, etc., as discussed in this paper). Sports machine learning models trained on motion-related data for understanding sports language can be trained on motion-related data to adjust one or more weights, layers, nodes, biases, and / or synapses. Sports machine learning models can include components (e.g., weights, layers, nodes, biases, and / or synapses) that collectively establish one or more of the following associations: player and team or league; team and player or league; score and team; scoring event and player; sporting event and player or team; winning and player or team; losing and player or team, etc. Sports machine learning models can establish associations between motion information and statistics within the context of a sporting event. Sports machine learning models can be trained to adjust one or more weights, layers, nodes, biases, and / or synapses to establish associations of certain motion statistics given the context of a sporting event. For example, a team's winning metric can be automatically associated with an opposing team's losing metric. As another example, a static score can be considered a positive attribution for the scoring team and a negative attribution for the losing team. As yet another example, a given score can be ranked relative to one or more other scores based on its relative position compared to those scores.

[0117] Sports machine learning models can be trained on sports tracking data and / or event data, as discussed in this paper. Such data can include player and / or object location information, movement information, trends, and changes. For example, a sports machine learning model can be trained by modifying one or more weights, layers, nodes, biases, and / or synapses to establish associations about the playing field and / or about the given location of one or more active subjects. As another example, a sports machine learning model can be trained by modifying one or more weights, layers, nodes, biases, and / or synapses to establish associations about the playing surface of the field and / or about the given movement or trend of one or more active subjects. As yet another example, a sports machine learning model can be trained by modifying one or more weights, layers, nodes, biases, and / or synapses to establish associations between sporting events and corresponding time boundaries, teams, players, coaches, referees, and environmental data associated with the location of the corresponding sporting event.

[0118] Sports machine learning models can be trained by modifying one or more weights, layers, nodes, biases, and / or synapses to establish associations of position, movement, and / or trend information in relation to motion objectives. Motion objectives can be score-related objectives (e.g., points, goals, shots, number of shots, scoring points, etc.), player outcomes (e.g., passes, objects such as ball movement, player positions, etc.), player positions, etc. Sports machine learning models can be trained in relation to motion objectives, match results, player positions, etc., associated with a given sport (e.g., football, American football, basketball, baseball, tennis, golf, rugby, hockey, team sports, individual sports, etc.). For example, a football-based sports machine learning model can be trained to establish associations or other connections about player positions on a football field. A football-based sports machine learning model can also be trained to establish associations or other connections about motion data involving multiple players and football-specific motion objectives.

[0119] Depending on various factors, the type of machine learning model for a given sport (e.g., generative learning, linear regression, logistic regression, random forest, gradient boosting machine (GBM), deep learning, graph neural network (GNN), and / or deep neural network) can be determined based on the attributes of the specific sport to which one or more machine learning models are applicable. Attributes may include, for example, sport type (individual sport vs. team sport), sport boundaries (e.g., time factors, number of players factors, object factors, possession periods (e.g., overlapping or independent), playing field type (e.g., restricted, unrestricted, virtual, real, etc.), player positions, etc.

[0120] Depending on the context, a sports machine learning model can receive input including motion data for a given sport and generate a matrix representation based on the features of that sport. The model can be trained to determine the latent features of the given sport. For example, the matrix may include fields and / or subfields related to player information, team information, object information, sport boundary information, sport surface information, etc. Attributes associated with each field and / or subfield can be populated within the matrix based on the received or extracted data. The model can perform operations based on the generated matrix. Features can be updated based on the input data or updated training data, which is based on, for example, motion data associated with features of a previously untrained model related to the given sport. Therefore, the model can be iteratively trained based on motion data or simulated data.

[0121] Retrieval-Enhanced Generation (RAG) is the process of optimizing the output of a large language model by referencing authoritative knowledge bases outside its training data sources before generating a response. Accessible data for general-purpose assistive tools has knowledge cutoff dates and needs to be continuously updated to include more information. Assistive tools, such as chatbots, can use LLM techniques to understand questions. For example, when a user enters a query, the chatbot will look for accessible data to respond, and if there is insufficient information within the accessible dataset, the chatbot can utilize RAG. By doing so, the chatbot can use the internet to augment its knowledge base and find answers. Using authenticated or verified websites, chatbots can utilize information found on the website to provide a response.

[0122] In such an implementation, the website can ingest up-to-date data and summarize user-relevant information, and the chatbot can utilize a subset of that data (e.g., summarized or modified data). However, this approach requires indexing and storing information that may exceed some LLM constraints (e.g., predefined context length). Furthermore, this approach may not be suitable for "live-time" events (e.g., events occurring at precise seconds during a match). This is likely due to potential latency issues caused by the website lacking available information at the time of the event.

[0123] One or more embodiments may utilize one or more of the following: match statistics, textual insights, predictions (e.g., match-level teams and players, and season-level teams), graphics, video overlays, and player and ball tracking data. For example, tracking data may be generated using in-field feeds or broadcast feeds. Tracking data may be supplemented with event data, which may be provided by an operator or automated system based on events related to a given sport within the field or via broadcast feeds. Tracking data and / or event data can be used to generate insights such as match statistics, textual insights, predictions, graphics, video overlays, etc. Therefore, the tracking data and insights generated according to the subject matter disclosed herein may be specific to a given sporting event and / or the sport associated with that event.

[0124] Retrieval-enhanced generation (RAG) methods (e.g., additionally or optionally utilizing a specific LLM trained for the motion, a fine-tuned LLM, or via cue engineering) can be implemented based on training one or more generative models (e.g., generative learning models or large language models). RAG methods can be implemented by first training a generative model using historical or simulated tracking data or insights associated with a given motion. The tracking data and / or insights may be relevant to the given motion and generated according to the techniques disclosed herein. Therefore, the generative model can be a motion-specific model for a given motion, such that it is optimized to generate the output for that given motion.

[0125] A production version of the generative model can be achieved by receiving a real-time data feed processed during a given competition. The real-time data feed may include tracking data and / or insights generated based on that tracking data. The real-time data feed can serve as a repository of information provided to the generative model. The real-time data feed can be provided in any format suitable for the generative model, such as, for example, a file containing data in key-value pairs, arrays, relations, matrices, etc. Alternatively or additionally, the generative model may convert the data associated with the real-time data feed into a format that can be applied by the generative model to generate its output.

[0126] Therefore, generative models can be trained using training data associated with a given sport and / or can be used by providing real-time data feeds for sporting events of that sport. Thus, generative models can be trained and used as specialized models, configured to generate nuanced outputs specifically for a given sport.

[0127] By using such a specialized generative model for a given sport, a user can ask questions in natural language (via input) and receive responses in natural language (e.g., via a display). Thus, a user can input queries and generate content such as images, video overlays, video animations, etc. For example, a user can use such a specialized generative model to input queries related to a given sport that are associated with the specialized generative model. The specialized generative model can generate outputs based on its training (e.g., using training data specific to the sport) and further based on one or more real-time data feeds associated with a match in that given sport. Based on both the training and the information available to the specialized generative model for the given sport, the specialized generative model can output a target output that is verifiable and meets an accuracy threshold unattainable by conventional techniques.

[0128] Figure 2 An exemplary flowchart for selecting agent-based artificial intelligence (AI) according to an example embodiment is depicted. (As mentioned above...) Figure 1The process 200 may be part of the orchestration module 124. Process 200 may include a proxy AI 210, a motion-specific proxy AI 220 (e.g., a motion-specific orchestrator as discussed herein), a language model 230, and a sports language 240. The proxy AI 210 may be referred to as the parent orchestrator 210. The parent orchestrator 210 may be configured to receive user input, including a description, and determine contextual and intent information related to the user input based on the description. The parent orchestrator 210 may determine the contextual and intent information by extracting one or more metadata items related to the description and determining at least one keyword or tag associated with the description. The parent orchestrator may also be configured to select one or more motion-specific orchestrators (e.g., 220). The parent orchestrator 210 may supervise the execution of actions or sets of actions related to the user input. For example, upon receiving user input (e.g., text, audio, video, drawing, etc.), the parent orchestrator 210 may determine the contextual and intent information associated with the description of the user input. The parent choreographer 210 does this to allow the selection of one or more sports-specific choreographers. For example, if the user's input involves Lionel Messi and Inter Miami, the parent choreographer can determine that the user is requesting information about a specific player "Lionel Messi," a team "Inter Miami," and the sport of football. The parent choreographer 210 can then select a football-specific choreographer based on the determined context and intent information.

[0129] The motion-specific agent AI 220 or motion-specific orchestrator 220 can be trained to select one or more agents to perform actions based on user input. The motion-specific orchestrator 220 can be trained using a motion-specific language. The motion-specific orchestrator 220 can be configured to select one or more agents based on the motion-specific language and corresponding attributes, as described in further detail below.

[0130] Language model 230 may include one or more sports-specific languages, each sports language being specific to a particular sport. Language model 230 may be used to train sports-specific choreographer 220. Sports language 240 may include one or more components, including tracking AI metrics 242, in-field tracking 244, remote tracking 246, event AI metrics 248, and event data 250. One or more components of sports language 240 may relate to the above-mentioned related... Figure 1One or more components of the computing environment 100. For example, tracking AI metric 242 may be substantially similar to tracking data system 116 and configured to generate and derive player tracking data based on broadcast feeds. On-field tracking 244 may be substantially similar to tracking system 102, which utilizes on-field devices to capture the movement of one or more active subjects (e.g., players) on the playing field, as well as the movement of one or more other related active subjects (e.g., objects) (e.g., ball, hockey, referee, etc.). Remote tracking 246 may be configured to capture player and ball tracking information from broadcast video (e.g., broadcast data). For example, remote tracking 246 may identify and track one or more players in the broadcast video of the current game / contest and map that information to real-world coordinates, e.g., on the field, court, or field, rather than within the broadcast video. Event AI metric 248 and event data 250 may be substantially similar to events and / or tracking as described herein. Each of one or more components of sports language 240 may determine a sports-specific language. For example, a sports-specific language may be based on sports-general attributes and sports-specific attributes. General sports attributes can be attributes applicable to more than one sport. For example, general sports attributes could include team versus individual, number of players, ground / field type, objects in the game, etc. Sports-specific attributes can be sports-specific, such as the type of lineup, whether the game is based on possession, how the game time is determined, what the available outcomes are, etc. Therefore, a sports-specific choreographer (e.g., a sports-specific agent AI 220) can be trained using a training dataset for that specific sport, allowing it to be trained based on the rules, context, and nuances of that specific sport.

[0131] For example, football can include one or more sport-specific attributes, including but not limited to a starting lineup with substitutes, player positioning based, play based on possession, continuous timing without interruption or delay, free movement (e.g., no stopping during possession), a goal being assigned a score of 1, a match result being win, loss, or draw / draw, certain actions (e.g., passing, fouling, shooting, corner kick, throw-in, etc.), and strategic concepts (e.g., formations, 4-4-2, counter-attack, etc.). Each of these sport-specific attributes can be used when determining events and / or tracking data (e.g., tracking AI metrics 242) from on-field tracking 244 and remote tracking 246. A sport-specific language model (e.g., football) can be generated based on training on the sport-specific attributes of that sport, as discussed herein. A sport-specific orchestrator 220 can be trained using a single sport-specific language. The sport-specific orchestrator 220 can then be configured to determine and select one or more agents for performing actions or sets of actions based on descriptions input by the user.

[0132] Figure 3A and Figure 3BAI agents according to an example embodiment are depicted. Figures 300A and 300B depict one or more agents and their associated connections. Each agent can receive and transmit information to each other to retrieve and perform actions or sets of actions based on user input descriptions. Each agent can be selected by a motion-specific orchestrator 220 based on determinations by a superior orchestrator 210. The superior orchestrator 210 can supervise the motion-specific orchestrator 220 and each of the selected agents performing actions or sets of actions. One or more agents can collaborate to collect, aggregate, and generate information and outputs related to the actions or sets of actions determined by the superior orchestrator 210 and the motion-specific orchestrator 220. According to embodiments disclosed herein, a first set of agents can be associated with a first motion and can be activated by a first motion-specific orchestrator for that first motion. The first set of agents can be trained using a training dataset of the first motion, such that they are trained based on rules, context, nuances, and actions associated with the first motion. Similarly, a second set of agents can be associated with a second motion and can be activated by a second motion-specific orchestrator for that second motion. The second group of agents can be trained using the training dataset of the second movement, allowing them to be trained based on the rules, context, nuances, and actions associated with the second movement.

[0133] One or more intelligent agents may include a visual question-answering agent 302, a player style comparison agent 304, a graph generator 306, a transfer market value predictor 308, a history researcher 310, a narrative generator 312, an insight agent 314, a player prop predictor 316, a match result predictor 318, a highlight generator 320, a real-time data agent 322, a ghost simulator 324, a player movement simulator 326, a set-piece strategy analyst 328, a formation predictor 330, a strategy adjuster 332, a referee analyst 334, a player development planner 336, a match integrity agent 338, an opponent weakness identifyer 340, and a motion data capture agent 342.

[0134] A visual question-answering agent 302 can be configured to provide one or more images, videos, match depictions, or other visual descriptions in response to a question or query. The visual question-answering agent 302 can use computer vision to determine one or more images and / or videos to answer a user's question. A player style comparison 304 can be configured to determine and provide a comparison of playing styles between one or more players and / or teams. For example, a player could be a defensive player rather than an offensive player. A graph generator 306 can be configured to determine and provide one or more graphs based on user input, such as a graphical representation of a team's wins compared to another team over a period of time. A transfer market value predictor 308 can be configured to determine the value of a single player being transferred or moved to another team. The transfer market value predictor 308 can be implemented using one or more prediction techniques disclosed herein. The transfer market value predictor 308 can provide indications about the player's potential value, monetary, or other aspects relative to the team. A historical researcher 310 can be configured to determine and collect historical information related to the description input by the user. The historical information can relate to specific points in time or ranges related to one or more players and / or teams. Narrative generator 312 can be configured to generate narratives or stories associated with one or more players and / or teams corresponding to user input. Stories or multiple stories can be generated based on historical, current, and / or predictive information. Narrative generator 312 can use one or more large language models to generate narratives. Insight agent 314 can be configured to generate one or more insights (e.g., average goals per game, playing time, wins per game, etc.) for one or more players and / or teams. Player prop predictor 316 can be configured to determine predictions for one or more players and their associated market odds. Predictions can utilize one or more of the player's current popularity, recent match statistics, etc. Match result predictor 318 can be configured to determine and predict the outcome of a match or contest between two teams or players. Match result predictor 318 can also generate predictions for series or tournaments between one or more teams. Highlights generator 320 can be configured to generate one or more graphics and text associated with event data for one or more players or teams. For example, highlights of Michael Jordan's rookie season.

[0135] Real-time data agent 322 can be configured to retrieve real-time data information related to one or more players, teams, and / or matches. Real-time data can come from real-time broadcast data from an on-field tracking system. Ghost simulator 324 can be configured to generate simulations of one or more scenarios and associated outcomes related to the match. For example, ghost simulator 324 can simulate the position and / or movement of players and / or objects to generate tracking data when players and / or objects are not visible in the feed used to generate tracking data (e.g., occluded or not presented in a frame). As another example, ghost simulator 324 can provide simulated match or tracking data based on proposed alternative inputs (e.g., simulations of modified matches, player positions, players, etc.). Player movement simulator 326 can be configured to generate simulations of a player's movement in a match relative to another player or team using one or more player events and / or tracking data. For example, simulating how Lionel Messi would move in a match against David Beckham. Set-piece strategy analyst 328 can be configured to determine specific tactical or strategic approaches for each restart of a match based on players, teams, positions, scores, etc. The formation predictor 330 can be configured to predict a team's formation during specific parts of a game. For example, the formation predictor 330 can determine whether a team's next down is likely to use a shotgun formation or a center-back formation when they are down by seven points, have two minutes left in the first half, and are at the opponent's 20-yard line. The strategy adjuster 332 can be configured to determine a set of strategies, tactics, or formations to be implemented during the game based on one or more criteria. Strategies, tactics, and / or formations can be dynamically updated based on historical, current, and predictive information. The referee analyzer 334 can be configured to analyze the influence of referees on the current competition or game, such as whether referees are biased towards one or more players from a particular team. The player development planner 336 can be configured to assess a player's ability to determine their success in a particular league, with a particular team, or against a particular player. The game integrity agent 338 can be configured to determine the integrity of the game or competition (e.g., fairness, transparency, or ethical behavior). The opponent weakness identifyr 340 can be configured to identify one or more players, strategies, or formations that provide weaknesses for a team. Additionally, the opponent weakness identifyer 340 can identify one or more weaknesses of a specific player (e.g., strength, speed, agility, etc.). The motion data capture agent 342 can be configured to retrieve and aggregate motion data (e.g., tactics, highlights, etc.) of a specific player or team.

[0136] Figure 4A flowchart depicts a method 400 for performing actions using agent-based artificial intelligence (AI) according to an example embodiment. In step 410, the method may include receiving one or more user inputs, wherein the user inputs include a description. The user inputs may include one or more of text input, drawing input, video input, audio input, event stream input, etc. The following is related to... Figure 5 and Figure 6 The user input is described in further detail. The description may include details and information related to players, teams, leagues, sporting events, etc. In step 420, the method may include determining contextual and intent information associated with the description. Contextual information may include relevant details that help clarify the situation regarding the query, the game (e.g., score, time, possession, momentum shifts, etc.), teams, players (e.g., name, position, history, etc.), player performance (e.g., statistics, injuries, fatigue, form history, etc.), or strategic decisions (e.g., lineup, coach, decisions, chemistry, etc.). Intent information may include data and / or insights reflecting the intentional actions, objectives, decisions, or strategies behind the query, player, or team. Intent information may also include purpose-driven information, decision-based information, or strategic or planning information. Determining contextual and intent information may include extracting one or more metadata items associated with the description and determining at least one keyword or tag associated with the description. One or more metadata items may include information that assists the higher-level choreographer 210 in determining one or more sports-specific choreographers 220 and corresponding multiple agents (e.g., as determined by the sports-specific choreographer 220). At least one keyword or tag associated with the description may include words or phrases that can indicate a specific player, team, event, or action requested by a user based on the description.

[0137] In step 430, the method may include selecting one or more sports-specific orchestrators (e.g., sports-specific orchestrator 220) based on multiple contextual and intent information associated with the description. For example, a user queries "provide highlights of Lionel Messi for Inter Miami in the last two seasons". The parent orchestrator may receive the input and determine the associated contextual and intent information. The contextual information may include "highlights" of a specific player "Lionel Messi" for a specific team "Inter Miami" during a specific time period "last two seasons". The intent information may include how to present the highlights to the user, such as a series of tactics, goals, victories, etc. In addition to the user profile, the intent information may be determined based on the description entered by the user, which may include past queries, output preferences, modifications, etc. This additional information may assist the parent orchestrator in determining the action or set of actions to be performed relative to the description entered by the user.

[0138] In step 440, the method may include selecting multiple agents for actions associated with the description based on contextual information and intent information. Selecting multiple agents may include mapping one or more metadata items to at least one or more event streams; and determining at least one keyword or tag associated with one or more metadata items, which are related to at least one or more event streams. (As mentioned above...) Figure 3A and Figure 3B Each of the multiple agents discussed can be configured to perform one or more actions. Each of the multiple agents can cooperate to collect, aggregate, or generate information requested by the user. A higher-level orchestrator and / or a motion-specific orchestrator can utilize one or more metadata items, event streams, keywords, or tags to determine which of the multiple agents have been trained to perform the requested specific action (e.g., individually or in combination with two or more agents). In one embodiment, the higher-level orchestrator and / or the motion-specific orchestrator can determine the multiple agents based on general motion attributes and motion-specific attributes as discussed herein. The higher-level orchestrator and / or the motion-specific orchestrator can determine the multiple agents based on a threshold of a motion-specific attribute, which is determined based on keywords, tags, metadata items, etc. When the threshold of the motion-specific attribute for a specific motion language is exceeded, the motion-specific orchestrator can select multiple agents corresponding to that specific motion.

[0139] In step 450, the method may include retrieving multiple pieces of information based on contextual and intent information. After the parent orchestrator selects one or more motion-specific orchestrators, each motion-specific orchestrator can then select multiple agents to perform actions or sets of actions related to user input. Each agent can be trained to retrieve specific information based on the actions assigned to it, such as one or more motion tracking data and one or more motion event data.

[0140] In step 460, the method may include determining an action to be performed in relation to the description input by the user. Performing the action may include generating one or more motion content items based on received information, and one or more agents may be configured to execute instructions to match at least one keyword or tag associated with the description with at least one keyword or tag associated with at least one or more event streams. Generating one or more motion content items may include retrieving one or more content items associated with a subset of at least one or more event streams. For example, a parent orchestrator and / or a motion-specific orchestrator may determine the type of content indicated from context and intent information (e.g., description, video, summary, etc.). Based on these decisions, the orchestrator may select appropriate actions to be performed and corresponding agents to perform each action.

[0141] In another example, the parent choreographer and / or the motion-specific choreographer can determine the request to perform match analysis based on context and intent information to determine the potential impact of the referee on the match outcome. The choreographer can select the appropriate action to be performed (e.g., historical matches and results over a period of time, including specific referee and / or team / player combinations) and the corresponding agent to perform each action (e.g., referee analyzer 334).

[0142] In yet another example, the choreographer and / or motion-specific choreographer can determine the request to make a match prediction based on context and intent information. The choreographer can select the appropriate action to be performed (e.g., retrieving historical match data for each team, player comparisons for each team, contextual results based on historical match information, etc.) and the corresponding agent to perform each action (e.g., match result predictor 318).

[0143] In another example, the choreographer and / or motion-specific choreographer can determine the request to predict the transfer market value of a specific player based on context and intent information. The choreographer can select the appropriate actions to be performed (e.g., identifying a player, determining the current composition of a team, the team ranking with / without a player, player influence, etc.) and the corresponding agent to perform each action (e.g., transfer market value predictor 308).

[0144] In step 470, the method may include activating an agent process to perform an action based on determined steps for performing the action. For example, using the example above, a user query for “provide highlights of Lionel Messi’s last two seasons at Inter Miami” can determine a set of actions and steps for performing that set of actions. Actions may include retrieving information related to Lionel Messi, in the form of tactics, goals, victories, etc. This information may be further defined as “highlights” to distinguish it from ordinary tactics or sporting events involving Lionel Messi. This information may be limited to a specific time period, “last two seasons,” and a specific club, “Inter Miami.” Furthermore, steps for performing the action may include aggregating the retrieved information and generating a montage or highlights based on that information (e.g., sports content items). Each agent selected by the sports-specific choreographer may then perform each action determined by the superior choreographer and / or the sports-specific choreographer. After completing the action, multiple agents may provide the sports content items to the superior choreographer for review. The superior choreographer may also request additional information and actions based on the sports content items presented by the multiple agents. The higher-level programmable controller can also provide motion content items to the user's device. The presentation of motion content items can be seen below related to... Figure 7 and Figure 8 Further detailed description.

[0145] In another example, based on a request regarding how current players will impact the league, the higher-level choreographer may have already determined multiple actions to be performed by one or more agents. The agents can then, for example, retrieve historical player information, player statistics, current and projected market value, coachability information, etc. Each agent can perform one or more actions based on requests from the higher-level choreographer and / or the sport-specific choreographer. Each agent can collect and aggregate information and generate output for review by the higher-level choreographer.

[0146] In yet another example, based on a request to identify the weaknesses of an opponent in an upcoming match, the higher-level choreographer may have already determined multiple actions to be performed by one or more agents. The agents can then, for example, retrieve player / team statistics, results of each match between players / teams, playing style information for each player / team, etc. Each agent can perform one or more actions based on requests from the higher-level choreographer and / or the sport-specific choreographer. Each agent can collect and aggregate information and generate output for review by the higher-level choreographer.

[0147] Figure 5 A flowchart depicts a method 500 for generating an interactive display for a sports content data stream, based on one aspect of the disclosed subject matter. In step 562, user input, including a description (e.g., a query), may be received. The description may be received via text input, audio input, video input, drawing input, gesture input, etc. The user input may be a description request related to a sports event, team, player, etc. In step 564, contextual information associated with the description may be determined. As discussed herein, a generative machine learning model (e.g., a parent orchestrator 210) may receive the description as input and may output contextual information based on the description. The contextual information may be related to the sports event data and / or content generated based on the sports event data.

[0148] In step 566, sports event and tracking data can be received and stored in a database. The sports event and tracking data can be automatically generated based on broadcast or in-field feeds and can include player and / or object location information, movement information, trends, changes, etc. As discussed herein, a motion-specific choreographer (e.g., motion-specific choreographer 220) can determine and select multiple agents to perform one or more actions associated with a description based on contextual and intent information. In one embodiment, multiple agents can retrieve sports event and tracking data stored in a database.

[0149] In step 568, multiple pieces of sports event and tracking data content (e.g., graphics, statistics, displays, images, videos, etc.) can be generated based on the sports event and tracking data received in step 566. Furthermore, advertisements and / or odds-based content can be generated based on user preferences and / or the sports event and tracking data received in step 566. As discussed herein, multiple agents can be configured to generate sports content items incorporating the retrieved sports event and tracking data.

[0150] In step 570, sports event and tracking data content can be filtered by matching such content to contextual information associated with the user input received in step 562. Matching can be performed by a machine learning model (e.g., multiple agents) that receives sports event and tracking data content, contextual information, and / or user preference information as input. In step 570, a subset of the content generated in step 568 can be output to the user as one or more sports content items. Sports content items can include one or more of the following: sports information cards, sports media stories, sports comparison graphs, or sports media that the user can passively consume or interact with (e.g., access additional content, place bets, access web pages or applications, etc.).

[0151] Figure 6 An example flowchart depicts a method 600 for generating tracking and / or event data based on one aspect of the disclosed subject matter. In step 602, one or more user inputs (e.g., user queries) may be received. The one or more inputs may include descriptions of sporting actions (e.g., tactics) and may be in text, audio, visualization, event / tracking data, etc., as discussed herein (e.g., references). Figure 7 and Figure 8 ).

[0152] In step 604, one or more metadata items associated with the description may be extracted. These metadata items can associate various aspects of the description with features that can be mapped to an event stream. The event stream can be historical or simulated tracking and / or event data determined based on broadcast or in-field motion streams. Therefore, in step 604, the user input query (e.g., the description) can be converted into a format that allows the input query to be mapped to an event stream. As discussed herein, a higher-level orchestrator (e.g., higher-level orchestrator 210) can receive user input and determine contextual and intent information by extracting one or more metadata items associated with the description of the user input. The contextual and intent information can assist the higher-level orchestrator in determining which of one or more motion-specific orchestrators might be needed to execute the action set. The contextual and intent information can provide additional details about potential outputs and / or actions included in the user input.

[0153] In step 606, metadata items can be mapped to one or more event streams, as referenced. Figure 8 Further discussion is needed. Mapped event streams and / or contextual information can be provided to the multimodal sports LLM model. In some embodiments, mapping metadata to event streams can be performed by a higher-level orchestrator. The mapped information can help determine which of the motion-specific orchestrators might be needed to perform one or more actions included in the user input.

[0154] In step 608, one or more content items associated with one or more mapped event streams can be output by the multimodal sports LLM. As discussed herein, the multimodal sports LLM can be trained to output actual or generated event and / or tracking data associated with the event stream mapped in step 606. The mapped event stream can provide features, criteria, and / or boundaries for information requested by the querying user in a format that allows the multimodal sports LLM to output responses to queries. In one embodiment, multiple agents selected by a motion-specific orchestrator can retrieve and generate output corresponding to actual or generated event and / or tracking data from the event stream.

[0155] In step 608, one or more content items output by the Multimodal Sports LLM may include actual or generated event and / or tracking data in response to a user query. The actual or generated event and / or tracking data may include player and / or object location information, movement information, trends, changes, etc., in response to a user query. Additionally, the output content items may include one or more actions based on the actual or generated event and / or tracking data.

[0156] In step 610, actual or generated event and / or tracking data output by the multimodal sports LLM can be provided to the user (e.g., via a user device). The output can be provided as a visual display depicting player and / or object location information, movement information, trends, changes, etc., in response to a user query. For example, player and / or object information can be provided in a video format depicting match actions corresponding to the player and / or object information. The video can run from the start to the end of a match action and may include indicators representing the player and / or object information. As another example, player and / or object information can be provided in an image format. The image can depict player and / or object information during a given match action. In one embodiment, multiple agents can perform one or more actions associated with metadata items related to a description entered by the user. For example, multiple agents can set market odds, update user streams and / or interfaces, record upcoming matches, purchase tickets, make reservations, etc.

[0157] In one embodiment, a user query might involve providing a turn-by-turn analysis of the current match. The system can retrieve one or more current broadcast streams (e.g., via real-time data agent 322) and provide a textual description of the match content (e.g., via narrative generator 312). Additionally, the turn-by-turn description could include predictions of each touch of the ball by one or more players (e.g., via match result predictor 318, player style comparison 304, player movement simulator 326, etc.), such as whether the current match action will result in a pass, a shot, or loss of possession, based on the current match situation.

[0158] In one embodiment, a user query may involve predicting the outcome of a current match. The system may receive one or more current broadcast streams (e.g., via real-time data agent 322) and / or historical match information (e.g., via historical researcher 310 and player style comparison 304). The system may also generate output (e.g., via graph generator 306 or visualization question-answering agent 302) based on the first half predictions (e.g., via insight agent 314, match result predictor 318, set-piece strategy analyst 328, opponent weakness identifyer 340, etc.) to display to the user.

[0159] Figure 7 and Figure 8 A flowchart is provided illustrating an exemplary method for performing actions using a machine learning model based on one aspect of the disclosed subject. Figure 7 An exemplary method 700 using a machine learning model is described. Method 700 begins by receiving user input from a client device (705-725). The input may take the form of a text description 705, an event stream 710, a drawing 715, audio 720, or video 725. For example, client device 108 may execute application 130 and provide a user interface for inputting one or more inputs. The user interface may include dialog boxes for inputting one or more of text, images, video clips, audio segments, etc., via drag and drop. Furthermore, the dialog boxes may include interface objects for capturing real-time audio and / or video from the user. Additional input may be received throughout the process to further refine the query.

[0160] Based on the received input, the system (e.g., orchestration module 124 and / or higher-level orchestrator 210) can determine context and intent information to describe the requested motion information. Context and / or intent information can be defined upon receiving the input, for example, defined by the user along with the query, and / or inferred by one or more models. For example, a user can input a query or question using one or more inputs as described above, and the query may include context and / or intent information. Orchestration module 124 can be configured to use one or more mapping models based on the input type to determine context and intent information based on the received input. Each mapping model used can be trained for a specific input type; for example, a drawing interface can be mapped to text questions, event streams, tracking data, images, videos, etc. The drawing mapping model can be configured to receive user input in a first format (e.g., drawing) and convert the received input into a second format usable by the multimodal sports LLM (e.g., game file 126). Input received by the multimodal sports LLM can be configured to infer context and / or intent information based on the query input. If context and / or intent information is received as part of the query, the multimodal sports LLM can also compare the input context and / or intent information with the inferred context and / or intent information to determine the accuracy of the multimodal sports LLM. The multimodal sports LLM can use the comparison of context and / or intent information to modify the training of each additional multimodal sports LLM to generate more accurate inferences for future inputs. Furthermore, the system may include a dialogue layer configured to clarify or request additional information from the user to determine the context and / or intent information relevant to the input query.

[0161] Based on context and intent information, method 700 can use a multimodal sports LLM 730 to determine sports event and / or tracking data that match the received input. As described herein, the multimodal sports LLM 730 may be part of orchestration module 124. Orchestration module 124, via a higher-level orchestrator 210, can determine one or more connections and / or associations with one or more sports event and / or tracking data based on context and intent information. Sports event and / or tracking data may include current and / or historical information associated with individual players and / or teams. After determining one or more sports event and / or tracking data, the system can output the sports event and / or tracking data to a user's client device (735) (e.g., client device 108) via multiple agents. The system may present sports tracking data and events in one or more of the following formats: audio and / or text commentary 740, visualization 745, similar tactics 750, and tactical analysis 755. For example, displaying highlights of Manchester United based on user queries (e.g., text or audio) could restrict highlights to goals against Liverpool. The mapping module 124 can determine the context and intent information associated with the query, and determine the connection and association with the motion tracking data. After retrieving motion tracking data from the tracking data system 116 and / or the data storage device 118, the system can output one or more clips of each goal scored in the Manchester United vs. Liverpool match. Furthermore, the user can enter further modifiers or queries to refine the context and intent information, further narrowing down the selected motion tracking data.

[0162] Figure 8 Alternative techniques for outputting tracking and event data are described. Figure 8 An exemplary method 800 using a machine learning model is described. (As with...) Figure 7 Similar to the description above, it can receive user input (e.g., 805, 810, 815, 820, and 825). Based on the received input, a preprocessing procedure can be performed (e.g., 860-870). Preprocessing can enhance the received input by determining what motion tracking data events were requested via the user query. For example, the system can determine the context of the query entered by the user during the preprocessing procedure. In addition to the context, the system can identify the intent behind the query, such as the question posed by the user. Using this additional information, the system can better map or match motion tracking data using multimodal sports LLM (830). As described above, the system then outputs motion tracking data and events (835) and presents the motion tracking data and events to the user in various ways (e.g., 840, 845, 850, and 855).

[0163] Preprocessing (860-870) may include mapping text and / or audio input to an event stream (e.g., using database lookup or other recognition processes). Therefore, during this preprocessing, the text and / or audio input is converted into a tracking and / or event data format, which can be provided to the multimodal sports LLM for more effective association with historical event and / or tracking data to generate output event and / or tracking data. (See above regarding...) Figure 7 The orchestration module 124, via the higher-level orchestrator 210, can perform some or all of the functions associated with preprocessing (860-870). Similarly, preprocessing (860-870) may include mapping visual data (e.g., plotting, video data, tracking data, etc.) to an event stream (e.g., using lookup or other recognition processes). The visual data can be analyzed and converted into a tracking and / or event data format, which can be provided to a multimodal sports LLM for more effective association with historical event and / or tracking data to generate output event and / or tracking data.

[0164] like Figure 8 The input data (825) described as being provided as an event stream may not require preprocessing, as such data may already be in a format that can be provided to a multimodal sports LLM in order to be more effectively correlated with historical events and / or tracking data to generate output events and / or tracking data.

[0165] Figure 9 Example scenarios for use in AR / VR and mixed reality applications (e.g., using head-mounted devices) are depicted. AR / VR and / or mixed reality devices can be used to receive user input and / or provide the user with tracking and / or event data, as described herein.

[0166] Figure 10 Example scenarios for use are described, in which a user can utilize a personal computer or TV as an input / output device. For example, a user can use AR / VR or other electronic devices to provide user input queries (e.g., such as...). Figure 7 and Figure 8 (as shown), for use by multimodal sports LLM to generate output tracking and / or event data.

[0167] Figure 11AAn exemplary input user interface according to one or more embodiments is depicted. User interface 1100A may include a competition dropdown menu 1110 configured to receive a user's selection of a competition (e.g., sport, league, tournament, etc.). Based on the competition selected via competition dropdown menu 1110, competition dropdown menu 1120 may be configured to populate based on competitions associated with the selected competition. A user may select one or more competitions via competition dropdown menu 1120, and one or more real-time data feeds associated with the selected one or more competitions may be provided to selective generative models, as discussed herein. User interface 1100A may also include an input box 1140 configured to allow a user to input one or more prompts for use by the one or more generative models. The user may utilize one or more inputs (e.g., text, audio, drawing, etc.), as referenced above. Figure 7 and Figure 8 The user interface 1100A may include a side panel 1150 configured to allow a user to create one or more new topics, enabling the clearing of queries or conversations associated with previously searched topics. The side panel 1150 may include a toggle button 1130 configured to show or hide the side panel 1150. Content displayed in the side panel 1150 may include a list of recent topics for selection, which may be automatically populated with the corresponding competition and / or match information selected from the competition dropdown menu 1110 and / or the competition dropdown menu 1120.

[0168] Figure 11B A user interface 1100B is depicted that responds to a user selecting a competition and a match via competition drop-down menus 1110 and 1120. Furthermore, the side panel 1150 has been hidden by a toggle button 1130. The user interface 1100B may include one or more sample query suggestions to assist in query generation. For example, sample query suggestions may be customized based on the selected competition, associated sport, and / or one or more selected matches. Additionally, sample query suggestions may be generated based on user profiles or user-related historical information, such as past suggestions.

[0169] Traditional generative AI applications fall short in providing information relevant to real-time sports. This is due to limited access to real-time data sources and subsequent augmenting data sources, such as insights, graphs, and predictions. The specialized generative model discussed in this paper overcomes these shortcomings by specifically training and / or using real-time data feeds (e.g., based on tracking data and insights) and enabling the specialized generative model to generate outputs based on this.

[0170] Overcoming the shortcomings of traditional generative AI applications includes using real-time data feeds being processed during the competition. Using this RAG-based approach, the input data can be "digitized," allowing specialized generative models to leverage this digitized data to respond to input queries.

[0171] According to embodiments, the dedicated generative model discussed herein can generate output based on user queries. User queries can be text queries, audio queries, queries generated based on user profiles, queries selected from a set of queries, etc. Outputs can be text output, audio output, images, graphics, videos, and / or interactive output. The dedicated generative model can generate multiple outputs and / or can use a first output to generate a second output. For example, the first output may include a text description, summary, or raw data in response to a user query. Such a first output can be provided to the user and / or used to generate a second output. The second output can be, for example, audio output, images, graphics, videos, and / or interactive output. Such a second output can be generated based on the first output data and the tracking data and / or insights associated with the real-time data feed upon which the first output is based. Furthermore, content from in-field feeds and / or broadcast feeds can be associated with the first output, such that the second output includes data associated with the first output as well as content extracted from in-field feeds and / or broadcast feeds.

[0172] For the generation of content (e.g., images, graphics, videos, and / or interactive output), necessary data can be extracted (e.g., from real-time data feeds) and subsequently fed to a graphics API to generate content. One advantage may include using real-time feeds to generate tracking data and / or analytics to enable RAG-based generative models. One or more embodiments may include the use of, for example, RAG and / or LangChain. One or more embodiments may include using a query local data storage device (which uses text2SQL constrained to the real-time match of interest).

[0173] One or more embodiments may include training a generative model using historical or simulated football data. The football generative model can adjust weights, layers, biases, synapses, etc., based on events, patterns, relationships, etc., associated with football. For example, the football generative model can be trained with boundaries including two teams per match, up to eleven players per team, one goalpost per team, one goalkeeper per team, one ball, a pass (a ball moving from one player to another), a goal (a ball moving into a goalpost), the meaning of red or yellow cards, stoppage time, etc.

[0174] The football generative model can receive football-specific real-time data feeds processed during one or more matches. These feeds can be provided by a RAG or LangChain system associated with them. Alternatively, the football generative model can query a local data storage device (e.g., using text2SQL constrained to real-time or applicable to the matches of interest). Information associated with the real-time data feeds can include, for example, player positions, ball position, referee position, number of players on the field for each team, remaining time, penalty kicks given and to which player, and remaining time on the penalty kick.

[0175] One or more embodiments may include a user inputting a football-related query (e.g., spoken, text, touch, etc., or a combination thereof) into a prompt. The input football-related query may relate to players, teams, matches / contests, or combinations thereof that the user is interested in. Once input, the query can be used to select one or more available football-specific real-time data feeds. Such real-time data feeds and football-specific training associated with a football generative model can be used to analyze the query, as discussed herein. Once analyzed, a response to the query can be generated, and output (e.g., text, audio, video, or a combination thereof) can be provided to the user. The output may be supplemented with content obtained from data storage devices and / or visual feeds (e.g., broadcast feeds and / or in-field feeds).

[0176] As a concrete example, a user can query about Lionel Messi and his CONCACAF Champions League live match against Monterrey via voice input. Further prompts can be provided to assist the user with additional details of interest. For example, a prompt could be made regarding how additional information should be displayed to the user in the output. One or more live data feeds can be provided to a football generative model. Such live data feeds can be used by a football-specific model to analyze information related to Lionel Messi, Inter Miami, the CONCACAF Champions League, each player and their corresponding position on the field, score, match events, etc. Live data feeds from the on-field system can track each player and / or team identified by the user query (e.g., Lionel Messi) using player-specific information (e.g., position, playing time, goals, shots, passes, etc.). The football generative model can generate outputs in response to queries based on its training and such live data feeds. Furthermore, responses can be provided to a graphics API that can generate visual content associated with Lionel Messi (e.g., an image of Lionel Messi identified as a high-engagement image in the live data feed). The output from the graph API can be in the form of visualizations, analyses, audio and / or text narration, or a combination thereof, based on the input query and / or additional prompts. For example, the initial query identified the player Lionel Messi, and the additional prompt identified a graph visualization of Lionel Messi with corresponding statistics throughout the match.

[0177] Figure 12 An example dialogue facilitated by a football generative model according to one or more embodiments is depicted. User interface 1200 can be configured to display one or more user prompts and responses based on a selected competition and match within chat window 1210. Data feeds can be based on applicable real-time data feeds and can include one or more depictions of user queries and responses generated using the football generative model discussed herein.

[0178] One or more embodiments may include a recommendation engine embedded in a chat window (e.g., chat window 1210) to assist the user. For example, the recommendation engine may learn from user behavior over time. The recommendation engine may automatically select one or more real-time data feeds based on the user's past queries. For example, if the user previously entered a query related to a specific team, the recommendation engine may update a dedicated generative model with real-time data from new matches associated with that team.

[0179] Figure 13An example graphical output interface according to one or more embodiments is depicted. User interface 1300 can provide an exemplary graphical representation 1310 generated in response to user queries and applicable real-time data feeds, as described above with reference to Figure 11 and... Figure 12 The graph 1310 may correspond to one or more queries input by the user. Exemplary graph 1310 may include one or more of text, graphics, audio, video, etc., to convey information requested by the user through one or more queries. For example, the user may input text or audio as a user query. Based on information determined by the system, a football generative model, as discussed herein, may be selected. The user interface 1300 may display graph 1310 depicting analytical data corresponding to football players as the subject of the user query. The analytical data may be generated based on one or more real-time data feeds and football-specific training of the football generative model.

[0180] Figure 14 A flowchart illustrating a method for training a machine learning model based on one aspect of the disclosed topic is provided. Figure 14 As shown in flowchart 1400, training data 1412 may include one or more of stage inputs 1414 and known results 1418 associated with the machine learning model to be trained. Stage inputs 1414 may originate from any applicable source, including the components or sets shown in the figures provided herein. Known results 1418 may be included for machine learning models generated based on supervised or semi-supervised training. Supervised machine learning models may be trained without using known results 1418. Known results 1418 may include known or expected outputs for future inputs that are similar to or belong to the same category as stage inputs 1414 that do not have corresponding known outputs.

[0181] Training data 1412 and training algorithm 1420 can be provided to training component 1430, which can apply training data 1412 to training algorithm 1420 to generate a trained machine learning model 1450. According to one embodiment, a comparison result 1416 can be provided to training component 1430, which compares the previous output of the corresponding machine learning model to apply the previous result to retrain the machine learning model. Training component 1430 can use comparison result 1416 to update the corresponding machine learning model. Training algorithm 1420 can utilize machine learning networks and / or models, including but not limited to, deep learning networks such as deep neural networks (DNN), convolutional neural networks (CNN), fully convolutional networks (FCN), and recurrent neural networks (RCN), probabilistic models (such as Bayesian networks and graphical models), and / or discriminative models (such as decision forests and maximum margin methods). The output of flowchart 1400 can be the trained machine learning model 1450.

[0182] The machine learning models disclosed herein can be trained by adjusting one or more weights, layers, and / or biases during the training phase. During the training phase, historical or simulated data can be provided as input to the model. The model can adjust one or more of its weights, layers, and / or biases based on such historical or simulated information. The adjusted weights, layers, and / or biases can be configured based on the training to a production version of the machine learning model (e.g., a trained model). Once trained, the machine learning model can output its output according to the subject matter disclosed herein. According to one implementation, one or more machine learning models disclosed herein can be continuously updated based on the use or implementation of the machine learning model's output.

[0183] Figure 15A An architecture of a computing system 1500 according to an example embodiment is shown. System 1500 may represent at least a portion of an organizational computing system 104. One or more components of system 1500 may communicate electrically with each other using bus 1505. System 1500 may include a processing unit (CPU or processor) 1510 and a system bus 1505 that couples various system components, including system memory 1515, such as read-only memory (ROM) 1520 and random access memory (RAM) 1525, to processor 1510. System 1500 may include a cache of high-speed memory that is directly connected to, close to, or integrated as part of processor 1510. System 1500 may copy data from memory 1515 and / or storage device 1530 to cache 1512 for fast access by processor 1510. In this way, cache 1512 can provide performance improvements by avoiding latency for processor 1510 while waiting for data. These and other modules can control or be configured to control processor 1510 to perform various actions. Other system memory 1515 may also be available. Memory 1515 may include various different types of memory with different performance characteristics. Processor 1510 may include any general-purpose processor and hardware or software modules, such as service 1 1532, service 2 1534, and service 3 1536 stored in storage device 1530, configured to control processor 1510; and dedicated processors, where software instructions are incorporated into the actual processor design. Processor 1510 may essentially be a completely independent computing system containing multiple cores or processors, buses, memory controllers, caches, etc. Multi-core processors may be symmetric or asymmetric.

[0184] To enable users to interact with the computing system 1500, the input device 1545 can represent any number of input mechanisms, such as a microphone for voice, a touchscreen for gesture or graphical input, a keyboard, a mouse, motion input, voice input, etc. The output device 1535 (e.g., a display) can also be one or more of a variety of output mechanisms known to those skilled in the art. In some cases, a multimodal system can enable users to provide multiple types of input to communicate with the computing system 1500. The communication interface 1540 typically governs and manages user input and system output. There are no limitations on operation on any particular hardware arrangement, and therefore the basic features described herein can be easily replaced with improved hardware or firmware arrangements during development.

[0185] Storage device 1530 may be non-volatile memory and may be a hard disk or other type of computer-readable medium that can store computer-accessible data, such as magnetic tape cassettes, flash memory cards, solid-state storage devices, digital multifunction optical discs, cartridge storage media, random access memory (RAM) 1525, read-only memory (ROM) 1520 and mixtures thereof.

[0186] Storage device 1530 may include services 1532, 1534, and 1536 for controlling processor 1510. Other hardware or software modules are contemplated. Storage device 1530 may be connected to system bus 1505. In one aspect, a hardware module performing a specific function may include software components stored in a computer-readable medium connected to necessary hardware components, such as processor 1510, bus 1505, output device 1535, etc., to perform that function.

[0187] Figure 15BA computer system 1550 is shown having a chipset architecture that can represent at least a portion of an organizational computing system 104. The computer system 1550 can be an example of computer hardware, software, and firmware that can be used to implement the disclosed techniques. System 1550 may include a processor 1555, which represents any number of physically and / or logically distinct resources capable of executing software, firmware, and hardware configured to perform the identified computations. The processor 1555 can communicate with a chipset 1560, which can control inputs to and outputs from the processor 1555. In this example, the chipset 1560 outputs information to an output 1565, such as a display, and can read and write information to a storage device 1570, which may include, for example, magnetic media and solid-state media. The chipset 1560 can also read data from and write data to RAM 1575. A bridge 1580 for interfacing with various user interface components 1585 may be provided for interfacing with the chipset 1560. Such a user interface component 1585 may include a keyboard, microphone, touch detection and processing circuitry, and indicating devices such as a mouse. Typically, input to system 1550 can come from any of a variety of machine-generated and / or human-generated sources.

[0188] Chipset 1560 can also interface with one or more communication interfaces 1590 that may have different physical interfaces. Such communication interfaces may include interfaces for wired and wireless local area networks (LANs), broadband wireless networks, and personal area networks (PANs). Some applications of the methods for generating, displaying, and using the GUI disclosed herein may include receiving ordered datasets via physical interfaces or analyzing data stored in storage device 1570 or RAM 1575 via processor 1555 to generate the data itself. Furthermore, the machine may receive input from the user via user interface component 1585 and perform appropriate functions, such as performing browsing functions by parsing this input using processor 1555.

[0189] It is understood that the example systems 1500 and 1550 may have more than one processor 1510 or be part of a group or cluster of computing devices networked together to provide greater processing power.

[0190] While the foregoing describes embodiments herein, other and additional embodiments can be devised without departing from the basic scope of the invention. For example, aspects of the invention can be implemented in hardware or software or a combination of hardware and software. One embodiment described herein can be implemented as a program product for use with a computer system. The program product defines the functionality of the embodiment (including the methods described herein) and can be contained on various computer-readable storage media. Illustrative computer-readable storage media include, but are not limited to: (i) non-writable storage media (e.g., read-only memory (ROM) devices within a computer, such as those that may be CD-ROM drives, flash memory, ROM chips, or any type of solid-state non-volatile memory) on which information is permanently stored; and (ii) writable storage media on which changeable information is stored (e.g., a floppy disk in a floppy disk drive or hard disk drive or any type of solid-state random access memory). Such computer-readable storage media are embodiments of the invention when they carry computer-readable instructions directing the functionality of the disclosed embodiments.

[0191] Those skilled in the art will understand that the foregoing examples are exemplary and not restrictive. It will be apparent to those skilled in the art, upon reading the specification and studying the accompanying drawings, that all permutations, enhancements, equivalents, and modifications are included within the true spirit and scope of the invention. Therefore, the appended claims are intended to cover all such modifications, permutations, and equivalents that fall within the true spirit and scope of these teachings.

Claims

1. A system for performing actions using agent-based artificial intelligence (AI), the system comprising: One or more motion-specific choreographers; Multiple agents, each agent being associated with at least one motion-specific orchestrator; A higher-level orchestrator, trained to select one or more motion-specific orchestrators, wherein the motion-specific orchestrators are trained to select one or more agents to perform actions based on one or more user inputs, wherein the higher-level orchestrator is configured to execute instructions for: Receive one or more user inputs, wherein the one or more user inputs include a description; Determine the contextual and intent information associated with the description; The one or more motion-specific orchestrators are selected based on the contextual and intent information associated with the description, wherein each of the one or more motion-specific orchestrators is trained using one or more motion-specific languages. The one or more motion-specific orchestrators are configured to execute instructions for the following: Based on the context information and the intent information, the plurality of agents are selected to perform the action associated with the description. The one or more intelligent agents are configured to execute instructions for the following: Retrieve multiple pieces of information based on the context information and intent information; Determine the steps for performing the action associated with the description; and The agent process is activated to perform the action based on the determined steps for performing the action.

2. The system of claim 1, wherein the one or more user inputs include at least one of text, audio, drawing, or video.

3. The system according to claim 1, wherein determining the plurality of context information and intent information further includes: Extract one or more metadata items related to the description; as well as Identify at least one keyword or tag associated with the description.

4. The system according to claim 1, wherein selecting the plurality of intelligent agents further comprises: Map one or more metadata items to at least one or more event streams; as well as Identify at least one keyword or tag associated with the one or more metadata items, which are related to the one or more event streams.

5. The system of claim 4, wherein performing the action includes generating one or more motion content based on the received information, and wherein the one or more agents are configured to execute instructions for: The at least one keyword or tag associated with the description will be matched with the at least one keyword or tag determined to be associated with the at least one or more event streams.

6. The system of claim 5, wherein generating the one or more motion content further includes retrieving one or more content items associated with a subset of the at least one or more event streams.

7. The system of claim 1, wherein each of the plurality of agents is trained to retrieve the plurality of pieces of information relating to one or more motion tracking data and one or more sports event data, and to arrange the one or more motion tracking data and the one or more sports event data for performing the action.

8. A method for performing motion-related actions, the method comprising: Receive one or more user inputs, wherein the one or more user inputs include at least a description; Identify multiple pieces of contextual and intent information associated with the description; One or more motion-specific language models are determined based on the aforementioned contextual and intent information; Based on the aforementioned contextual and intent information, one or more sports events and tracking data are determined; as well as Use the one or more sports-specific language models to retrieve one or more content items related to the one or more sports events and tracking data; as well as Transmit the one or more content items for display on the user's device.

9. The method of claim 8, wherein the one or more motion-specific languages ​​are trained using one or more general motion attributes and motion-specific attributes.

10. The method of claim 9, wherein the general sports attributes include the number of players, field type, team sport, or individual sport.

11. The method of claim 9, wherein the sport-specific attribute includes starting lineup, ball-handling-based sport, segmented sport, time limit, score distribution, or free-throw-based sport.

12. The method of claim 9, wherein determining the plurality of context and intent information further comprises: Extract one or more metadata items related to the description; Identify at least one keyword or tag associated with the description; as well as Map the one or more metadata items to the at least one keyword or tag.

13. The method of claim 12, wherein determining one or more motion-specific language models further comprises: Match the at least one keyword or tag with the general attributes of the sport and the specific attributes of the sport; The threshold number is determined for each identified motion-specific attribute in the one or more motion-specific language models. If the number of motion-specific attributes identified for one of the one or more motion-specific language models exceeds the threshold number, then one of the one or more motion-specific languages ​​is selected.

14. A non-transitory computer-readable medium configured to store processor-readable instructions, wherein when executed by a processor, the instructions perform the following operations: A higher-level orchestrator, trained to select one or more motion-specific orchestrators, wherein the motion-specific orchestrators are trained to select multiple agents to perform actions based on one or more user inputs, wherein the higher-level orchestrator is configured to execute instructions for: Receive one or more user inputs, wherein the one or more user inputs include a description; Determine the contextual and intent information associated with the description; The one or more motion-specific orchestrators are selected based on the contextual and intent information associated with the description, wherein each of the one or more motion-specific orchestrators is trained using one or more motion-specific languages. The one or more motion-specific orchestrators are configured to execute instructions for the following: Based on the context information and the intent information, the plurality of agents are selected to perform the action associated with the description. The one or more intelligent agents are configured to execute instructions for the following: Retrieve multiple pieces of information based on the context information and intent information; Determine the steps for performing the action associated with the description; and The agent process is activated to perform the action based on the determined steps for performing the action.

15. The non-transitory computer-readable medium of claim 14, wherein the one or more user inputs include at least one of text, audio, drawing, or video.

16. The non-transitory computer-readable medium of claim 14, wherein determining the plurality of context information and intent information further comprises: Extract one or more metadata items related to the description; as well as Identify at least one keyword or tag associated with the description.

17. The non-transitory computer-readable medium of claim 14, wherein determining the plurality of intelligent agents further comprises: Map one or more metadata items to at least one or more event streams; as well as Identify at least one keyword or tag associated with the one or more metadata items, which are related to the one or more event streams.

18. The non-transitory computer-readable medium of claim 17, wherein performing the action includes generating one or more motion content based on the received information, and wherein the one or more agents are configured to execute instructions for: The at least one keyword or tag associated with the description will be matched with the at least one keyword or tag determined to be associated with the at least one or more event streams.

19. The non-transitory computer-readable medium of claim 18, wherein generating the one or more motion content further includes retrieving one or more content items associated with a subset of the at least one or more event streams.

20. The non-transitory computer-readable medium of claim 14, wherein each of the plurality of agents is trained to retrieve the plurality of pieces of information relating to one or more motion tracking data and one or more sports event data, and to arrange the one or more motion tracking data and the one or more sports event data for performing the action.

Images