Rapid game making method based on artificial intelligence and game making cloud platform

Abstract

The invention relates to a quick game making method and a game making cloud platform, and the method comprises the steps: binding received natural language demand description data, sketch track data and real-time voice streams into an associated data object which can be stored for a long time, and distributing a global unique ID and a version updating record identifier for the data object; according to the response emergency degree of the instruction in the data object, distributing a differentiated data transmission channel; performing time point matching on the voice features, the sketch structure features and the text semantic features of the same data object through a space-time synchronization engine, and outputting a unified instruction vector; calling a pre-trained game content generation model, and converting the unified instruction vector into game elements which can be independently stored; and monitoring the modification operation of a player on the game elements, extracting modification difference data, training and generating a model optimization module, and updating a version record. According to the method, the effects of cross-session state inheritance of the NPC / scene, dynamic preemptive scheduling of computing power resources and reverse correction of the AI model by the player operation data are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of game production, and in particular to a game rapid production method based on artificial intelligence and a game production cloud platform. BACKGROUND

[0002] With the vigorous development of the electronic game industry and the increasing diversification and high frequency of user demand for game content, game development efficiency and cost control have become the core competitive elements in the industry. In recent years, the breakthrough of artificial intelligence (AI) technology has provided automated tools for game development, and some tools and platforms aimed at accelerating game development have appeared in the market, such as template-based rapid development frameworks, visual script editing tools, and online game material libraries.

[0003] However, the current AI game production platform can automatically generate basic content such as scenes and characters, but there are essential bottlenecks, such as NPC behavior and scene parameters that cannot be persisted across sessions, and players need to reinitialize each interaction; cannot dynamically schedule according to real-time operations, with a latency of over 2000ms in high concurrency; player correction operations only trigger local recompilation, without reverse optimization of the underlying model, leading to repeated errors, etc.

[0004] Based on the above-mentioned defects, it is urgent to break through the traditional technical bottlenecks and build a new generation of game production engine that supports state persistence, resource elastic allocation, and reverse model optimization. SUMMARY

[0005] To solve at least one of the above technical problems, the present application provides a game rapid production method based on artificial intelligence and a game production cloud platform.

[0006] In a first aspect, the present application provides a game rapid production method based on artificial intelligence, which adopts the following technical solution: The following steps are performed on a cloud server: Bind the received natural language requirement description data, sketch trajectory data, and real-time voice stream into an associated data object that can be stored for a long time, and assign a globally unique ID and a version update record identifier to the data object; According to the response urgency of the instructions in the data object, assign a differentiated data transmission channel; Match the voice features, sketch structure features, and text semantic features of the same data object at different time points through a space-time synchronization engine, and output a unified instruction vector; Call a pre-trained game content generation model to convert the unified instruction vector into a game element that can be independently stored, the game element including an NPC element, a scene structure element, and a task flow element; The modification operation of the player on the game element is monitored, modification difference data is extracted, a generation model optimization module is trained, and a version record is updated.

[0007] By adopting the technical solution, the cross-session inheritance of the NPC / scene is realized by associating the data object and the version record, such as the NPC remembering the tactical preference of the player yesterday, thereby solving the state persistence problem; meanwhile, the differentiated transmission channel dynamically allocates bandwidth according to the emergency degree, thereby breaking the resource allocation rigidity; in addition, the modification difference data is closed-loop to the model optimization module, thereby greatly reducing the same error and realizing the reverse model optimization. Therefore, the cross-session state inheritance of the NPC / scene, the dynamic preemption scheduling of the computing resource, and the effect of the reverse correction of the AI model by the player operation data can be realized.

[0008] In a possible implementation manner, the step of allocating the differentiated data transmission channel according to the response emergency degree of the instruction in the data object specifically includes: The voice stream is divided into semantic paragraphs, and is associated to the key anchor point of the sketch track through sound feature matching; The text description is subjected to syntax structure analysis, a core action word is extracted, and is mapped to the center area of the sketch structure.

[0009] By adopting the technical solution, the multi-modal instruction association is utilized to eliminate the multi-source signal timing mismatch, and the fusion error rate is greatly reduced.

[0010] In a possible implementation manner, in the step of calling the pre-trained game content generation model to convert the unified instruction vector into the game element that can be independently stored, the NPC element construction includes: Player behavior preference features are extracted from historical interaction records; A behavior attenuation coefficient is set, the initial value of which is 1.0, and the coefficient is attenuated by 0.1 every 24 hours without interaction, and the NPC initiatively interacts when the coefficient is less than 0.5.

[0011] By adopting the technical solution, the NPC behavior entity construction is utilized to realize the continuous evolution of the NPC behavior.

[0012] In a possible implementation manner, the step of monitoring the modification operation of the player on the game element, extracting the modification difference data, training the generation model optimization module, and updating the version record includes: When the player modifies the NPC dialogue logic, the semantic difference data before and after the modification is recorded; The difference data is input into the training process containing the adversarial mechanism, and the dialogue logic discrimination ability of the generation model is strengthened.

[0013] By adopting the technical solution, the dialogue logic reverse optimization is utilized to solve the problem of repeated occurrence of dialogue errors.

[0014] In a possible implementation manner, the method further includes: When the player logs in again, loading the latest version record of the data object according to the global ID; Automatically playing back the key change points in the version update for the player to confirm.

[0015] By using the above technical solution, the cross-session state synchronization is used to ensure that the player inherits the complete game state each time the player logs in.

[0016] In a second aspect, the application provides a game production cloud platform system, which includes: The instruction reasoning cluster includes parallel voice processors, sketch analyzers and text analyzers, and is connected through a dynamic bandwidth bus, and the bus adjusts the data transmission bandwidth between the modules in real time according to the data type; The element repository stores game elements and version records thereof, and the game elements include a basic information area, a behavior record area and a structure change area; The resource scheduler has a built-in real-time load detector, dynamically reallocates GPU resources according to the number of players and the operation type, and the resource allocation priority strategy is voice processing>real-time battle>scene generation; The feedback learning gateway includes a player operation difference extractor and a model fine-tuning adapter, and converts the player modification operation into model training data.

[0017] By using the above technical solution, the instruction processing cluster is used, and the parallel processors are connected through the dynamic bus to solve the problem of signal blocking; the resource scheduler is used, and real-time load detection effectively reduces the voice processing delay compared with static allocation; the feedback learning gateway is used, and the player operation directly drives the model fine-tuning to reduce the cost of manual debugging.

[0018] In a possible implementation manner, the dynamic allocation strategy of the resource scheduler is: When the voice interaction amount>threshold, the GPU resource weight of the scene generation module is reduced to 10%, and the voice processing module is increased to 70%; When the battle test request amount surges, a degradation rendering mode is enabled and 20% of the video memory is released for behavior logic processing.

[0019] By using the above technical solution, the resource utilization rate of the high-concurrency scene is effectively improved by using dynamic resource allocation.

[0020] In a possible implementation manner, the player operation difference extractor of the feedback learning gateway performs: Extracting movement trajectory optimization data from the operation of the player dragging the NPC path; Extracting difficulty change gradient from the operation record of the player adjusting the level difficulty.

[0021] By adopting the technical solution, the operation difference extraction is adopted to accurately capture the player's intention, and model directional optimization is realized. BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1 A flowchart of a game rapid production method based on artificial intelligence provided by an embodiment of the present application.

[0023] Figure 2 A structural diagram of a game production cloud platform provided by an embodiment of the present application.

[0024] Figure 3 A structural diagram of an electronic device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0025] The technical solutions in the present application will be described below with reference to all the drawings. Obviously, the described embodiments are only some of the embodiments of the present application, not all the embodiments.

[0026] In the description of the embodiments of the present application, unless otherwise specified, " / " represents the meaning of "or", for example, A / B can represent A or B; "and / or" in this document is only a description of the association relationship of the associated objects, which means that there can be three relationships, for example, A and / or B, which can represent: A exists alone, A and B exist together, and B exists alone. In addition, in the description of the embodiments of the present application, "plurality" or "multiple" means two or more than two.

[0027] Hereinafter, the terms "first" and "second" are only for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of indicated technical features. Therefore, the features defined with "first" and "second" can explicitly or implicitly include one or more features. In the description of the embodiments, unless otherwise specified, the meaning of "multiple" is two or more than two.

[0028] The terms used in the following embodiments are only for the purpose of describing specific embodiments and are not intended to be limiting to the present application. As used in the specification and the appended claims of the present application, the singular forms "a," "an," and "the" are intended to include, for example, the expression "one or more," unless the context clearly indicates otherwise. It should also be understood that in the following embodiments of the present application, "at least one" and "one or more" mean one, two or more than two.

[0029] Reference to "one embodiment" or "some embodiments" or "one implementation" or "some implementations" means that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment. The appearances of the phrase "in one embodiment" or "in some embodiments" in various places in the specification are not necessarily all referring to the same embodiment, although it can. The terms "including," "comprising," "having" and variations thereof are meant to encompass the items listed thereafter, but do not exclude other items from being present.

[0030] Embodiments of the present application provide a game rapid production method based on artificial intelligence, which is executed by an electronic device. The electronic device can be a stand-alone physical electronic device, an electronic device cluster or a distributed system composed of multiple physical electronic devices, or a cloud electronic device providing cloud computing services. Embodiments of the present application do not limit this, as shown in the following figure, the method comprises the following steps executed on a cloud server: Figure 1 S1, binding the received natural language requirement description data, sketch trajectory data, and real-time voice stream into an associated data object that can be stored for a long time, and assigning a globally unique ID and version update record identifier to the data object.

[0031] Specifically, the cloud server is equipped with three types of data receiving interfaces: natural language description is transmitted through a network port into a text buffer; hand-drawn sketch trajectory is stored in a trajectory buffer by digitizing tablet coordinate sequence collection; voice stream is converted into digital signal packets through a microphone array.

[0032] Further, the three types of data input objects are bound to a processor, which includes: Unique ID generator: using a hardware random source to generate a 128-bit identity code; Version recorder: creating a version chain start marker and change identification bit in the storage chip; Data association unit: merging and storing the three types of data through a high-speed electronic switch, and writing the identity code and version number in the header of the storage area.

[0033] Further, the bound data packet is stored in a non-volatile memory (such as a flash memory array) through a direct storage controller, and the storage location is recorded in an address mapping table. When a new data packet is ready, an interrupt signal is triggered to notify the channel allocation module to prepare for reception.

[0034] S2, according to the response urgency of the instruction in the data object, assign a differentiated data transmission channel.

[0035] Specifically, the data transmission channel allocation system includes three levels of control: ​Priority discrimination unit: voice data triggers red alarm signal, directly connects dedicated high-speed channel; sketch data triggers yellow alarm signal to enter buffer channel with flow control; text data goes through green ordinary channel.

[0036] Channel controller: voice channel assigns independent physical line to ensure transmission speed; sketch channel sets bandwidth guarantee mark; text channel adopts large-capacity transmission protocol.

[0037] Dynamic regulator: real-time monitoring of channel congestion, automatic compression of text channel bandwidth when voice delay exceeds standard.

[0038] Further, the voice data is divided into meaning paragraphs, matched to the key positioning points of the sketch through the voiceprint recognition chip; the text data is extracted through the grammar parser Core Verb, mapped to the center coordinate area of the sketch.

[0039] S3, through the space-time synchronization engine, the voice features, sketch structure features and text semantic features of the same data object are matched at the time point, and a unified instruction vector is output.

[0040] Specifically, the space-time synchronization engine includes: Clock reference source: high-precision clock chip provides synchronous time mark.

[0041] Feature processing unit: voice feature extractor generates voiceprint map; sketch analyzer outputs structure contour data; text processor generates word meaning code.

[0042] Alignment buffer: three types of feature data with time mark are stored in dual-port memory, and time window controller performs feature matching within 50ms tolerance range.

[0043] Further, the matched feature data is input into the splicing circuit to generate combined instruction data packet. When time misalignment occurs, the synchronization error counter triggers a re-matching instruction.

[0044] S4, calling the pre-trained game content generation model, converting the unified instruction vector into a game element that can be independently stored.

[0045] Among them, the game elements include NPC elements, scene structure elements, and task flow elements.

[0046] Specifically, the game content generation model includes: The combined instruction data packet is transmitted to the AI processing chip through the high-speed bus.

[0047] Three parallel working areas in the chip: role generation area, outputting role behavior parameters (including decay counter and preference record); scene generation area, generating scene coordinate data and collision detection template; task generation area, compiling task trigger mark and branch selection table.

[0048] The generated data is written into the element database by the storage controller: the basic information area, the creator's identity and the generation time are stored in the storage chip; the behavior record area, the interaction history is stored in the circular storage area; the structure record area, the scene change record is stored by tree index.

[0049] Further, the role behavior decay counter is automatically decremented daily, and when the value is lower than the threshold, the interaction event generation circuit is triggered.

[0050] S5, monitor the player's modification operation on the game element, extract the modification difference data, train the generated model optimization module, and update the version record.

[0051] Specifically, the player feedback optimization module includes: Operation monitoring circuit: player modification operation is converted into electronic operation code.

[0052] Difference extractor: dialogue modification, calculate the difference value before and after text modification; path adjustment, record the change amount of moving track.

[0053] Model updater: difference data is stored in training cache, triggering model update process for locking original model parameter storage area, loading difference data set, starting parameter update state machine.

[0054] Further, the version update adopts the "copy-on-write" technology, and the new version data is stored in the new storage area, and the version chain pointer is automatically jumped. When the player logs in again, the identity code directly locates the latest data, and the version comparator generates a change report to the display interface.

[0055] In summary, by using hardware-level data binding and non-volatile storage principles, permanent game state preservation (cross-session inheritance) is achieved; by using three-level alarm channel and dynamic bandwidth adjustment principles, voice command fast response (delay ≤200ms) is achieved; by using clock synchronization and dual-port storage matching principles, multi-source instruction precise fusion is achieved, error rate is reduced; by using three-way parallel generation and dedicated storage principles, NPC adaptive evolution is achieved; by using difference electronic calculation and copy-on-write principles, model intelligent optimization is achieved, error rate is reduced.

[0056] Based on this, the cross-session state inheritance of NPC / scene, the dynamic preemptive scheduling of computing power resources, and the effect of reverse correction of AI model by player operation data can be realized.

[0057] In some embodiments, in order to eliminate multi-source signal timing mismatch and achieve a greatly reduced fusion error rate, S2 specifically includes: S201, divide the voice stream into semantic paragraphs, and associate them to the key anchor points of the sketch track through sound feature matching.

[0058] Specifically, the voice stream is input into the voice segmentation module through the audio interface, which contains: Endpoint detection circuit: real-time monitoring of voice amplitude with energy threshold comparator, output segmentation pulse signal when 5ms continuous below the silence threshold.

[0059] Sketch segment generator: segmentation pulse triggers storage controller, stores voice stream into independent storage block of dual-port RAM.

[0060] Voiceprint matcher: pre-stored user voiceprint feature code lookup table (LUT) receives voice segment, outputs similarity value to comparator, and activates associated signal when matching is successful.

[0061] Further, the associated signal triggers the coordinate binding unit to write the storage block address of the current voice segment into the key anchor point coordinate register of the sketch trajectory.

[0062] Further, the anchor points are marked by the sketch preprocessing module, including trajectory inflection point coordinates (identified by curvature calculation circuit) and user manually marked points (captured by click event detector). The binding result is stored in the association relationship mapping table.

[0063] S202, perform syntax structure analysis on the text description, extract the core action word, and map it to the center area of the sketch structure.

[0064] Specifically, the text data is input into the syntax analysis unit, and the hardware executes the following process: Part-of-speech tagger: identify verbs / nouns through finite state machine, and mark verbs as action word candidates.

[0065] Main stem extractor: filter core verbs based on dependency rule library (stored in ROM), such as extracting "build" when receiving "build [object] at [location]"; extract "move" when receiving "let [character] move to [location]".

[0066] Coordinate mapper: input core verb code into address decoder, output corresponding sketch area coordinate range.

[0067] Further, the sketch structure division uses the area gridding processor to divide the bounding box of the sketch trajectory into 8x8 grids, and marks the high-density area as the center area through the edge density detection circuit. The matching result of the coordinate range of the core verb and the center area grid is stored in the center mapping register.

[0068] In summary, the hardware endpoint detection and voiceprint LUT matching principles are used to realize accurate association of voice instructions with spatial positions; finite state machine syntax analysis and grid partitioning are used to realize dynamic binding of text instructions to core areas.

[0069] In some embodiments, in order to realize the continuous evolution of NPC behavior, the NPC element construction in S4 includes: S401, extract player behavior preference features from historical interaction records.

[0070] Specifically, the historical interaction record memory (such as DDR4 memory) outputs the original operation code stream to the behavior feature extraction circuit, which includes, Operation classifier: identify the player operation type (attack / defense / collaboration) through the instruction decoder, and output the type label pulse.

[0071] Frequency counter: 24-hour cumulative count for each type of operation pulse.

[0072] Priority encoder: compare the operation count values of each type, and encode the highest frequency type as a 3-bit preference feature code.

[0073] Further, the preference feature code is written into the feature vector register group, such as attack type mapped to 001, defense type to 010, and collaboration type to 100. The register output end is connected to the mode selection pin of the NPC behavior generator, which controls the subsequent behavior logic generation path.

[0074] S402, set the behavior decay coefficient, its initial value is 1.0, decay 0.1 every 24 hours without interaction, and trigger the NPC active interaction behavior when the coefficient <0.5.

[0075] Specifically, the behavior decay system is composed of the following hardware: Real-time clock module: DS3231 chip is used to generate 24-hour timing pulse.

[0076] Decay control unit: contains an initial value register, which is written with an 8-bit binary value "11111111" (corresponding to a behavior decay coefficient of 1.0) when power-on reset; decrementer, arithmetic right shift (equivalent to ×0.9) is performed every time a timing pulse is received; threshold comparator, LM393 chip is used to compare the decay register output with the threshold "01111111" (set to 0.5).

[0077] Further, when the comparator output is low (behavior decay coefficient <0.5), the NPC interaction enable circuit is triggered, the power gate of the NPC dialogue generator is enabled, the behavior selector automatically loads the preset interaction script (stored in ROM), and the interaction request data packet is sent to the player terminal through the UART interface.

[0078] In summary, the effects of quantifying player behavior preferences, avoiding software count delays, NPC behavior evolving autonomously over time, and implementing NPC active interaction mechanisms are achieved.

[0079] In some embodiments, in order to solve the problem of repeated dialogue errors, S5 specifically includes: S501, when the player modifies the NPC dialogue logic, record the semantic difference data before and after modification.

[0080] Specifically, the player modification operation generates an operation code stream through an input device (such as a keyboard / handle), which is input into a dialogue difference recording module, which includes: A modification capture unit, the original dialogue text is stored in the A storage area of the dual-port RAM, and the modified text is written in the B storage area in real time.

[0081] A difference calculation circuit, the word comparator compares the contents of A / B area byte by byte, and the difference position marker writes the different byte address into the FIFO queue.

[0082] A semantic vector generator, the difference text segment is input into the word vector lookup table, and a 32-bit difference feature vector is output and stored in the difference data register.

[0083] Further, the recording process triggers a version marker circuit to create a new node in the version chain memory, and the node header stores the difference vector address pointer and the time stamp counter value.

[0084] S502, input the difference data into the training process containing the adversarial mechanism, and strengthen the dialogue logic discrimination ability of the generated model.

[0085] Specifically, the adversarial mechanism training hardware includes: A training data selector, channel 0 is the original generated dialogue vector (read from the NPC element storage area); channel 1 is the difference vector after the player modification (read from the difference register). Among them, the selection signal is output by the adversarial controller in a 1:1 ratio alternately.

[0086] It also includes a discrimination ability enhancer, which includes a two-way input comparator for comparing the similarity of channel 0 / 1 vectors; a weight update state machine for triggering parameter adjustment when the similarity is greater than the threshold, such as freezing the original model weight storage area, activating the negative sample training path, and adjusting the discrimination threshold voltage comparator.

[0087] It also includes a feedback loop for writing the result to the model discrimination flag register and updating the dialogue generation logic gate array through the system bus.

[0088] Further, the training process enables hardware acceleration, and the difference vector is directly transmitted to the training buffer through the DMA channel, and the clock frequency of the adversarial controller is increased to 3 times that of the basic mode.

[0089] In summary, the effects of accurately capturing dialogue modification points, avoiding real-time calculation delay, strengthening error recognition ability, and implementing adversarial reinforcement mechanism are achieved.

[0090] In some embodiments, to ensure that the player inherits the complete game state each time he logs in, the method further comprises: S6, when the player logs in again, load the latest version record of the data object according to the global ID.

[0091] Specifically, when the player logs in, the identity recognition code (such as a two-dimensional code / fingerprint) is input, and the global ID database is matched through the ID verification circuit, which includes: Address locator: input 128-bit global ID into operation chip (such as SHA-256 hardware accelerator), and output value as storage address index.

[0092] Version loading unit: index address is located in version chain head pointer in non-volatile memory through address decoder, and the latest version data packet (including NPC state + scene coordinates + task progress) is loaded into cache through DMA controller.

[0093] State recovery circuit: cache data is distributed to each processing module through system bus, NPC behavior register group is written with historical decay coefficient, and scene coordinate latch is loaded with the latest position data.

[0094] Further, the loading process enables the data verifier, and the cyclic redundancy check circuit verifies the integrity of the version data, and sends a ready signal to the game engine after the verification is passed.

[0095] S7, automatically replay the key change points in version update for the player to confirm.

[0096] Specifically, the version playback system includes: Change point extractor: version difference tree memory outputs node marker stream (using B+ tree hardware accelerator), and key node filter identifies nodes with a change amplitude greater than 20% (such as NPC behavior mutation points).

[0097] Playback data generation unit: in scene change, coordinate difference data is input into 3D rendering pipeline to generate scene evolution animation frames; in behavior change, state machine jump record drives NPC action generator to output behavior demonstration script.

[0098] Confirmation interface: animation frame sequence is stored in video frame buffer, and output to player display through HDMI controller, and confirmation button state is read into confirmation flag register through interface.

[0099] Further, when the confirmation signal is detected, the version synchronization circuit performs atomic write operation to write the playback node marker into the version confirmation record area.

[0100] The game production cloud platform provided by the embodiments of the present application is introduced below, and the game production cloud platform described below can be correspondingly referred to the game rapid production method based on artificial intelligence described above.

[0101] Reference Figure 2 , the game production cloud platform comprises: An instruction inference cluster 1, the instruction inference cluster 1 contains voice processors (including ADC sampling circuits), sketch analyzers (with coordinate track caches), text analyzers (with built-in word vector lookup tables) in parallel, and is connected through a dynamic bandwidth bus between modules. The bus adjusts the data transmission bandwidth between modules in real time according to the data type.

[0102] Specifically, the modules are interconnected through a dynamic bandwidth control bus, as follows: The bus controller monitors the data type identification code in real time, such as voice data triggering a high-priority flag (red signal light), sketch priority (yellow signal light), and text low priority (green signal light).

[0103] The electronic switch array switches the physical path according to the signal light state, such as 16 data channels (total bandwidth 80Gbps) in the red state, 8 data channels (40Gbps) in the yellow state, and 4 data channels (20Gbps) in the green state.

[0104] Among them, the flow monitor counts the data backlog of each module, and triggers a bandwidth multiplication signal when the cache exceeds the threshold.

[0105] Further, when the voice processor outputs the voiceprint feature package, it automatically raises the bus priority and occupies the text processor channel resources through an interrupt signal. The measured voice processing delay is stable at ≤200ms.

[0106] An element repository 2 is used to store game elements and their version records, and the game elements include a basic information area, a behavior record area, and a structure change area.

[0107] Specifically, in the basic information area, such as creator ID and timestamp are stored in a high-speed read-write chip; in the behavior record area, such as NPC interaction abstract uses circular storage technology, and the write pointer automatically circulates to cover the earliest data; in the structure change area, such as scene coordinate changes use version tree storage.

[0108] A resource scheduler 3 has a built-in real-time load detector, which dynamically reallocates GPU resources according to the number of players and operation types, and its resource allocation priority strategy is voice processing (default 60%) > real-time battle (default 25%) > scene generation (default 15%).

[0109] A feedback learning gateway 4 contains a player operation difference extractor and a model fine-tuning adapter, which converts player modification operations into model training data.

[0110] The instruction processing cluster 1, the parallel processor combined with the dynamic bus solve the problem of signal blocking; the resource scheduler 3, real-time load detection makes the voice processing delay effectively reduced compared with static allocation; the feedback learning gateway 4, player operation directly drives model fine tuning, reduces the cost of manual debugging.

[0111] In some embodiments, in order to effectively improve the resource utilization of high concurrency scenarios, the dynamic allocation strategy of the resource scheduler 3 is: When the voice interaction volume is greater than the threshold, the GPU resource weight of the scenario generation module is reduced to 10%, and the voice processing module is increased to 70%.

[0112] When the battle test request volume surges, the degraded rendering mode is enabled and 20% of the video memory is released for behavior logic processing.

[0113] In some embodiments, in order to accurately capture the player's intention and realize model directional optimization, the player operation difference extractor of the feedback learning gateway 4 executes: Extract the movement trajectory optimization data from the player's operation of dragging the NPC path.

[0114] Extract the difficulty change gradient from the player's operation record of adjusting the level difficulty.

[0115] The embodiments of the application provide an electronic device, such as Figure 3 as shown, Figure 3 a structural schematic diagram of an electronic device provided by the embodiments of the application, Figure 3 The electronic device 300 shown in the figure includes a processor 301 and a memory 303. Wherein, the processor 301 and the memory 303 are connected, such as connected through the bus 302. Optionally, the electronic device 300 can also include a transceiver 304. It should be noted that in actual application, the transceiver 304 is not limited to one, and the structure of the electronic device 300 does not constitute a limitation on the embodiments of the application.

[0116] The processor 301 can be a CPU (Central Processing Unit), a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array) or other programmable logic device, transistor logic device, hardware component, or any combination thereof. The processor 301 can implement or execute various exemplary logical blocks, modules and circuits described in combination with the disclosure of the embodiments of the present application. The processor 301 can also be a combination of computing functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, and the like.

[0117] The bus 302 can include a path for transmitting information between the above-mentioned components. The bus 302 can be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, or the like. The bus 302 can be divided into an address bus, a data bus, a control bus, and the like. For ease of representation, Figure 3 Only one thick line is used in the figure, but it does not mean that there is only one bus or only one type of bus.

[0118] The memory 303 can be a ROM (Read Only Memory) or other type of static storage device that can store static information and instructions, a RAM (Random Access Memory) or other type of dynamic storage device that can store information and instructions, an EEPROM (Electrically Erasable Programmable Read Only Memory), a CD-ROM (Compact Disc Read Only Memory) or other optical disk storage, a magnetic disk storage medium or other magnetic storage device, or any other medium that can be used to carry or store desired program code in the form of instructions or data structures and that can be accessed by a computer, but is not limited thereto.

[0119] The memory 303 is used to store application program code for implementing the embodiments of the present application, and is controlled by the processor 301 to execute. The processor 301 is used to execute the application program code stored in the memory 303 to realize the content shown in the foregoing method embodiments.

[0120] The electronic device includes, but is not limited to, a mobile terminal such as a mobile phone, a notebook computer, a digital broadcast receiver, a PDA (Personal Digital Assistant), a PAD (Tablet Personal Computer), a PMP (Portable Multimedia Player), a car terminal (for example, a car navigation terminal), and the like, and a stationary terminal such as a digital TV, a desktop computer, and the like. Figure 3 The illustrated electronic device is merely an example and should not impose any limitation on the function and use range of the embodiments of the present application.

[0121] The embodiments of the present application provide a computer readable storage medium, and the computer readable storage medium stores a computer program. The computer program is executed by a processor to implement the steps of the above-mentioned game rapid production method based on artificial intelligence.

[0122] Since the embodiments of the computer readable storage medium part correspond to the embodiments of the method part, the embodiments of the computer readable storage medium part are described with reference to the description of the embodiments of the method part.

[0123] It should be understood that, although each step in the flowchart of the accompanying drawings is displayed in sequence according to the direction of the arrow, these steps are not necessarily executed in sequence according to the direction of the arrow. Unless otherwise specified herein, the execution of these steps is not strictly limited in sequence, and they can be executed in other sequences. Moreover, at least part of the steps in the flowchart of the accompanying drawings can include multiple sub-steps or multiple stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution sequence is not necessarily sequential, but can be alternately or alternately executed with at least part of other steps or sub-steps or stages of other steps.

[0124] The above is only some embodiments of the present application, and it should be pointed out that, for those skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, and these improvements and refinements should be considered as the protection scope of the present application.

Claims

1. A method for rapid game production based on artificial intelligence, characterized in that, Perform the following steps on the cloud server: The received natural language requirement description data, sketch trajectory data, and real-time voice stream are bound into a long-term associated data object, and a globally unique ID and version update record identifier are assigned to the data object. Differentiated data transmission channels are allocated based on the urgency of the instructions in the data object; The spatiotemporal synchronization engine performs time-point matching of the speech features, sketch structure features, and text semantic features of the same data object, and outputs a unified instruction vector. A pre-trained game content generation model is invoked to transform a unified instruction vector into independently storable game elements, including NPC elements, scene structure elements, and task flow elements. Monitor player modifications to game elements, extract modification difference data, train and generate a model optimization module, and update version records.

2. The method according to claim 1, characterized in that, The step of allocating differentiated data transmission channels based on the urgency of the instructions in the data object specifically includes: The speech stream is segmented into semantic segments, and key anchor points of the sketch trajectory are associated with the speech feature matching. Perform grammatical structure analysis on the text description, extract core action words, and map them to the central area of ​​the sketch structure.

3. The method according to claim 1, characterized in that, In the step of calling the pre-trained game content generation model to transform the unified instruction vector into independently storable game elements, the NPC element construction includes: Extract player behavior preference features from historical interaction records; Set a behavior decay coefficient with an initial value of 1.

0. The decay is 0.1 every 24 hours without interaction. When the coefficient is less than 0.5, NPC active interaction behavior is triggered.

4. The method according to claim 1, characterized in that, The steps of monitoring player modifications to game elements, extracting modification difference data, training and generating a model optimization module, and updating version records include: When players modify NPC dialogue logic, the semantic differences before and after the modification are recorded; By inputting differential data into a training process that incorporates adversarial mechanisms, the ability of the generative model to discriminate dialogue logic is enhanced.

5. The method according to claim 1, characterized in that, Also includes: When a player logs in again, the latest version of the data object is loaded based on the global ID; Automatically replay key changes in the version update for players to confirm.

6. A game development cloud platform system, characterized in that, include: The instruction inference cluster includes parallel speech processors, sketch parsers, and text analyzers. The modules are connected through a dynamic bandwidth bus, which adjusts the data transmission bandwidth between modules in real time according to the data type. An element repository stores game elements and their version records. The game elements include a basic information area, a behavior record area, and a structural change area. The resource scheduler has a built-in real-time load detector that dynamically reallocates GPU resources based on the number of players and the type of operation. Its resource allocation priority strategy is: voice processing > real-time combat > scene generation. The feedback learning gateway includes a player action difference extractor and a model fine-tuning adapter, which convert player actions into model training data.

7. The system according to claim 6, characterized in that, The dynamic allocation strategy of the resource scheduler is as follows: When the amount of voice interaction exceeds the threshold, the GPU resource weight of the scene generation module will be reduced to 10%, while the weight of the voice processing module will be increased to 70%. When combat test requests surge, enable degraded rendering mode and free up 20% of video memory for behavioral logic processing.

8. The system according to claim 6, characterized in that, The player action difference extractor of the feedback learning gateway executes: Extract movement trajectory optimization data from player dragging NPC paths; Extract the difficulty change gradient from the player's operation record of adjusting the level difficulty.

Images