Moving track prediction method and related device

Abstract

The invention discloses a moving track prediction method and a related device, which are applied to scenes such as artificial intelligence and games. The method comprises the steps of representing object state information, game environment information and game task information and features associated with a target object in a first target time period in a target game as object state features, game environment features and game task features; performing hybrid coding to obtain a prediction output vector of the target object in a second target time period after the first target time period in the target game; and decomposing each prediction output vector into a prediction mean vector, a prediction standard deviation vector and a prediction probability vector so as to accurately determine a prediction moving track of the target object in the second target time period in the target game. Even if the movement track of the player in the game is very complex, the method can accurately predict the movement track of the player in the game in the future time period, so that the prediction effect of the movement track in the game is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer, in particular to a moving trajectory prediction method and related device. BACKGROUND

[0002] With the rapid development of artificial intelligence technology, the moving trajectory of a player in a future time period can be predicted by the moving state information of the player in a past time period in a game.

[0003] In the related art, the moving trajectory of the player in the future time period can be obtained by combining the moving state information of the player in the past time period with a physical model for predicting the moving trajectory.

[0004] However, the moving trajectory of the player in the game is very complex, and the physical model in the above method cannot accurately predict the moving trajectory of the player in the future time period, resulting in poor prediction effect of the moving trajectory in the game. SUMMARY

[0005] To solve the above technical problems, the present application provides a moving trajectory prediction method and related device, which can accurately predict the moving trajectory of the player in the future time period in the game even if the moving trajectory of the player in the game is very complex, thereby improving the prediction effect of the moving trajectory in the game.

[0006] The embodiments of the present application disclose the following technical solutions:

[0007] In one aspect, the present application provides a moving trajectory prediction method, comprising:

[0008] characterizing the object state information, the game environment information and the game task information associated with the target object in a first target time period in a target game to obtain object state features, game environment features and game task features; the object state information includes moving state information;

[0009] mixing encoding the object state features, the game environment features and the game task features to obtain a prediction output vector of the target object in a second target time period in the target game; the second target time period is after the first target time period;

[0010] For each prediction output vector, vector decomposition is performed on the prediction output vector to obtain a prediction mean vector, a prediction standard deviation vector and a prediction probability vector;

[0011] According to the prediction mean vector, the prediction standard deviation vector and the prediction probability vector, the prediction moving trajectory of the target object in the second target time period in the target game is determined.

[0012] In another aspect, an embodiment of the present application provides a mobile trajectory prediction device, the device comprising: a feature representation unit, a hybrid encoding unit, a vector decomposition unit, and a determination unit.

[0013] The feature representation unit is configured to perform feature representation on object state information, game environment information, and game task information associated with a target object in a first target time period of a target game, to obtain object state features, game environment features, and game task features; the object state information comprises mobile trajectory information.

[0014] The hybrid encoding unit is configured to perform hybrid encoding on the object state features, the game environment features, and the game task features, to obtain predicted output vectors of the target object in a second target time period of the target game; the second target time period is after the first target time period.

[0015] The vector decomposition unit is configured to, for each of the predicted output vectors, perform vector decomposition on the predicted output vector, to obtain a predicted mean vector, a predicted standard deviation vector, and a predicted probability vector.

[0016] The determination unit is configured to determine, according to the predicted mean vector, the predicted standard deviation vector, and the predicted probability vector, a predicted mobile trajectory of the target object in the second target time period of the target game.

[0017] In another aspect, an embodiment of the present application provides a computer device, the computer device comprising a processor and a memory:

[0018] The memory is configured to store a computer program and transmit the computer program to the processor;

[0019] The processor is configured to execute the method according to the instructions in the computer program.

[0020] In another aspect, an embodiment of the present application provides a computer readable storage medium for storing a computer program, when the computer program is run on a computer device, the computer program causes the computer device to execute the method according to any one of the preceding aspects.

[0021] In another aspect, an embodiment of the present application provides a computer program product comprising a computer program, when the computer program is run on a computer device, the computer program causes the computer device to execute the method according to any one of the preceding aspects.

[0022] It can be seen from the technical solution that first, the object state information, the game environment information and the game task information associated with the target object in the first target time period in the target game are represented as object state features, game environment features and game task features, wherein the object state information includes movement state information; and the object state features, the game environment features and the game task features are mixed and encoded as a prediction output vector of the target object in the second target time period after the first target time period in the target game. By converting the object state information, the game environment information and the game task information associated with each time step of the target object in the past time period in the target game into features suitable for the model, and capturing the association relationship between the object state information, the game environment information and the game task information in a mixed encoding manner, the movement features of the target object in each time step in the future time period in the target game, i.e., the prediction output vector, are predicted.

[0023] Then, for each prediction output vector, the prediction output vector is decomposed into a prediction mean vector, a prediction standard deviation vector and a prediction probability vector; and the prediction movement trajectory of the target object in the second target time period in the target game is determined based on the prediction mean vector, the prediction standard deviation vector and the prediction probability vector. By decomposing the prediction output vector into the prediction mean vector, the prediction standard deviation vector and the prediction probability vector, the movement position of the target object in each time step in the future time period in the target game can be predicted, so as to represent the movement trajectory of the target object in the future time period. Based on this, even if the movement trajectory of the player in the game is very complex, the movement trajectory of the player in the future time period in the game can be accurately predicted by the method, so as to improve the prediction effect of the movement trajectory in the game. BRIEF DESCRIPTION OF DRAWINGS

[0024] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0025] Figure 1 A system schematic diagram of a movement trajectory prediction method provided by an embodiment of the present application;

[0026] Figure 2 A flowchart of a movement trajectory prediction method provided by an embodiment of the present application;

[0027] Figure 3 A flowchart of a movement trajectory prediction method provided by an embodiment of the present application;

[0028] Figure 4A flowchart of training a mixed coding model and a Gaussian mixed model is provided for an embodiment of the present application.

[0029] Figure 5 A schematic diagram of downstream application of a mobile trajectory is provided for an embodiment of the present application.

[0030] Figure 6 A structural diagram of a mobile trajectory prediction device is provided for an embodiment of the present application.

[0031] Figure 7 A structural diagram of a server is provided for an embodiment of the present application.

[0032] Figure 8 A structural diagram of a terminal is provided for an embodiment of the present application. DETAILED DESCRIPTION

[0033] The embodiments of the present application will be described below with reference to the accompanying drawings.

[0034] At the present stage, the mobile trajectory of a player in a game is obtained by combining the mobile state information of the player in a past time period with a physical model used for predicting the mobile trajectory, and predicting the mobile trajectory of the player in a future time period. However, it is found through research that the mobile trajectory of a player in a game is very complex, and the physical model cannot accurately predict the mobile trajectory of the player in the future time period, resulting in poor prediction effect of the mobile trajectory in the game.

[0035] An embodiment of the present application provides a mobile trajectory prediction method, which converts object state information, game environment information and game task information associated with each time step of a target object in a past time period in a target game into features suitable for a model, and captures the association relationship between the object state information, the game environment information and the game task information in a mixed coding manner, so as to predict the mobile features of the target object in each time step in a future time period in the target game, i.e., a prediction output vector; by decomposing the prediction output vector into a prediction mean vector, a prediction standard deviation vector and a prediction probability vector, the mobile position of the target object in each time step in the future time period in the target game can be predicted, so as to represent the mobile trajectory of the target object in the future time period in the target game.

[0036] Next, the system architecture of the mobile trajectory prediction method will be introduced. Referring to Figure 1 , Figure 1 A system schematic diagram of a mobile trajectory prediction method is provided for an embodiment of the present application, which includes a computer device 100 configured to execute the mobile trajectory prediction method.

[0037] The computer device 100 performs feature representation on object state information, game environment information, and game task information associated with the target object in a first target time period in the target game, to obtain object state features, game environment features, and game task features; the object state information includes movement state information.

[0038] As an example, the target game is a shooting game, the first target time period is T1, T1 includes c time steps, c is a positive integer, c≥2, and the target object is a target player; then the computer device 100 performs feature representation on c object state information, c game environment information, and c game task information associated with the target player in T1 in the shooting game, to obtain c object state features o_11,…,o_1c, c game environment features g_11,…,g_1c, and c game task features m_11,…,m_1c; the object state information includes movement state information.

[0039] The computer device 100 performs hybrid encoding on the object state features, the game environment features, and the game task features, to obtain a predicted output vector of the target object in a second target time period in the target game; the second target time period is after the first target time period.

[0040] As an example, on the basis of the above example, the second target time period is T2, T2 is after T1, T2 includes d time steps, d is a positive integer, d≥2; then the computer device 100 performs hybrid encoding on o_11,…,o_1c, g_11,…,g_1c, and m_11,…,m_1c, to obtain d predicted output vectors y_21,…,y_2d of the target player in T2 in the shooting game.

[0041] The computer device 100 performs vector decomposition on each predicted output vector, to obtain a predicted mean vector, a predicted standard deviation vector, and a predicted probability vector.

[0042] As an example, on the basis of the above example, the computer device 100 performs vector decomposition on each y_t2, t2=21,…,2d, to obtain a predicted mean vector mean_t2, a predicted standard deviation vector scale_t2, and a predicted probability vector logit_t2.

[0043] The computer device 100 determines a predicted movement trajectory of the target object in the second target time period in the target game according to the predicted mean vector, the predicted standard deviation vector, and the predicted probability vector.

[0044] As an example, on the basis of the above example, the computer device 100 determines the predicted moving trajectory of the target player in T2 in the shooting game as y'_T2 = [y'_21, …, y'_2d] according to mean_t2, scale_t2 and logit_t2.

[0045] That is, by converting the object state information, game environment information and game task information of the target object in the target game associated with each time step in the past time period into features suitable for the model, and capturing the association relationship between the object state information, game environment information and game task information in a hybrid encoding manner, the moving feature of the target object in the target game at each time step in the future time period, i.e., the prediction output vector, is predicted; by decomposing the prediction output vector into a predicted mean vector, a predicted standard deviation vector and a predicted probability vector, the moving position of the target object in the target game at each time step in the future time period can be predicted, thereby representing the moving trajectory of the target object in the target game in the future time period; even if the moving trajectory of the player in the game is very complex, the moving trajectory of the player in the game in the future time period can be accurately predicted, thereby improving the prediction effect of the moving trajectory in the game.

[0046] It should be noted that the moving trajectory prediction method in the embodiments of the present application relates to artificial intelligence. Artificial intelligence is the use of digital computers or digital computer controlled machines to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technology of computer science, which tries to understand the essence of intelligence and produce a new intelligent machine that can react in a similar way to human intelligence. Artificial intelligence is to study the design principles and implementation methods of various intelligent machines, so that machines have the functions of perception, reasoning and decision-making.

[0047] Artificial intelligence technology is a comprehensive discipline, involving a wide range of fields, both hardware and software technologies. Artificial intelligence basic technologies generally include, such as sensors, special artificial intelligence chips, cloud computing, distributed storage, big data processing technology, pre-training model technology, operation / interaction system, mechatronics, etc. Among them, the pre-training model is also called large model, basic model, which can be widely used in downstream tasks of various directions of artificial intelligence after fine-tuning. Artificial intelligence software technology mainly includes computer vision technology, speech processing technology, natural language processing technology and machine learning / deep learning, etc. In the embodiments of the present application, artificial intelligence technology mainly involves machine learning / deep learning technology.

[0048] Machine learning / deep learning is a multi-disciplinary subject, involving probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory, etc. It is a specialized study of how computers simulate or implement human learning behavior to acquire new knowledge or skills, reorganize existing knowledge structure to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental approach to making computers intelligent. Its applications are widespread in various fields of artificial intelligence. Machine learning and deep learning usually include artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and teaching learning. Pre-trained models are the latest development of deep learning, which integrates the above technologies.

[0049] It should be noted that in the embodiments of the present application, the computer device can be a server or a terminal, and the method provided by the embodiments of the present application can be executed by the terminal or the server alone, or by the terminal and the server together. When the method provided by the embodiments of the present application is executed by the terminal or the server alone, the execution method is similar to Figure 1 The corresponding embodiments are similar, mainly replacing the computer device with a terminal or a server. In addition, when the method provided by the embodiments of the present application is executed by the terminal and the server together, the steps that need to be reflected on the front-end interface can be executed by the terminal, and some steps that do not need to be reflected on the front-end interface and need to be calculated in the background can be executed by the server.

[0050] The terminal can be a smartphone, a tablet computer, a notebook computer, a desktop computer, a smart voice interaction device, a vehicle-mounted terminal, an extended reality device, or an aircraft, but is not limited thereto. The server can be a standalone physical server, a server cluster or a distributed system composed of multiple physical servers, or a cloud server providing cloud computing services, but is not limited thereto. The terminal and the server can be directly or indirectly connected through wired or wireless communication, which is not limited in the present application. For example, the terminal and the server can be connected through a network, which can be a wired or wireless network.

[0051] In addition, the embodiments of the present application can be applied to various scenarios, including but not limited to cloud technology, artificial intelligence, intelligent transportation, assisted driving, autonomous driving, games, digital humans, virtual humans, virtual reality, augmented reality, mixed reality, audio and video, etc.

[0052] Next, the mobile trajectory prediction method provided by the embodiments of the present application will be described in detail with the computer device executing the method provided by the embodiments of the present application as an example and in conjunction with the drawings.

[0053] Referring to Figure 2 , Figure 2 The flowchart of the mobile trajectory prediction method provided by the embodiments of the present application, the method comprises:

[0054] S201: Feature representation is performed on object state information, game environment information and game task information associated with the target object in the first target time period in the target game, to obtain object state features, game environment features and game task features; the object state information includes movement state information.

[0055] S202: The object state features, the game environment features and the game task features are mixed coded to obtain a prediction output vector of the target object in the second target time period in the target game; the second target time period is after the first target time period.

[0056] In the related art, the movement trajectory of the target player in the future time period is obtained by combining the movement state information of the target player in the past time period in the target game with a physical model used for predicting the movement trajectory. However, it is found through research that the movement trajectory of the target player in the target game is very complex, and the physical model cannot accurately predict the movement trajectory of the target player in the future time period in the target game, resulting in poor prediction effect of the movement trajectory in the game.

[0057] Therefore, in the embodiments of the present application, considering that the movement trajectory of the target player in the target game is very complex, and the player state information, the game environment information and the game task information associated with the target player in the past time period in the target game affect the movement trajectory of the target player in the future time period, when the target player is set as the target object, the past time period is set as the first target time period, and the future time period is set as the second target time period, it is necessary to feature represent the object state information, the game environment information and the game task information associated with the target object in the first target time period in the target game as object state features, game environment features and game task features that can be processed by the model, wherein the object state information includes movement state information; and mix code the object state features, the game environment features and the game task features to capture the association relationship between the object state information, the game environment information and the game task information, so as to predict the prediction output vector of the target object in the second target time period in the target game, to represent the movement features of the target object in each time step in the second target time period in the target game.

[0058] The target game is a game that the game player needs to move; the first target time period refers to a past time period; the target object refers to a target player among the game players; one object state information corresponds to one time step in the first target time period, and refers to movement state information and life state information, etc., the movement state information refers to one or more of position information, speed information, acceleration information and orientation information, and the life state information refers to one or more of blood volume information and life value information; one game environment information corresponds to one time step in the first target time period; one game task information corresponds to one time step in the first target time period; one object state feature refers to a feature corresponding to the object state information in one time step in the first target time period; one game environment feature refers to a feature corresponding to the game environment information in one time step in the first target time period; one game task feature refers to a feature corresponding to the game task information in one time step in the first target time period; the second target time period refers to a future time period; and one prediction output vector corresponds to one time step in the second target time period and refers to a movement feature.

[0059] The S201-S202 converts the object state information, the game environment information and the game task information associated with the target object in the target game in the past time period into features suitable for the model, so that the model can process the object state information, the game environment information and the game task information; on this basis, the association relationship between the object state information, the game environment information and the game task information is captured in a hybrid coding manner, so as to predict the movement feature of the target object in the target game in each time step in the future time period, that is, the prediction output vector, and provide a data basis for subsequent prediction of the moving position of the target object in the target game in each time step in the future time period.

[0060] As an example of S201-S202, the target game is a shooting game, the first target time period is T1, T1 includes c time steps, c is a positive integer, c≥2, the target object is a target player, the second target time period is T2, T2 is after T1, T2 includes d time steps, d is a positive integer, d≥2; the computer device converts the c object state information, the c game environment information and the c game task information associated with the target player in the shooting game T1 into c object state features o_11, …, o_1c, c game environment features g_11, …, g_1c and c game task features m_11, …, m_1c, wherein the object state information includes movement state information; and o_11, …, o_1c, g_11, …, g_1c and m_11, …, m_1c are hybrid coded into d prediction output vectors y_21, …, y_2d in T2 in the shooting game.

[0061] The game environment information can be ray information, and the game environment feature can be a ring-shaped ray feature, that is, a ray is emitted in a direction in which the target player is facing, and a distance of a first ray collision point of each ray is recorded.

[0062] S203: For each prediction output vector, vector decomposition is performed on the prediction output vector to obtain a prediction mean vector, a prediction standard deviation vector, and a prediction probability vector.

[0063] S204: According to the prediction mean vector, the prediction standard deviation vector, and the prediction probability vector, a predicted movement trajectory of the target object in the second target time period in the target game is determined.

[0064] In the embodiments of the present application, after the prediction output vector of the target object in the second target time period in the target game is predicted by the object state information associated with the target object in the first target time period in the target game, the game environment information, and the game task information, it is considered that the movement trajectory of the target player in the target game is very complex, and in order to capture the complexity of the trajectory, for each prediction output vector, the prediction output vector is decomposed into parameters of a Gaussian model, that is, a prediction mean vector, a prediction standard deviation vector, and a prediction probability vector. The movement position of the target object at each time step in the second target time period in the target game can be predicted through the prediction mean vector, the prediction standard deviation vector, and the prediction probability vector, so as to determine the predicted movement trajectory of the target object in the second target time period in the target game.

[0065] The prediction mean vector includes a plurality of mean vectors corresponding to a plurality of Gaussian distributions of the movement position of the target object at one time step in the second target time period in the target game. The target prediction standard deviation vector includes a plurality of standard deviation vectors corresponding to a plurality of Gaussian distributions of the movement position of the target object at one time step in the second target time period in the target game. The target prediction probability vector includes a plurality of weights corresponding to a plurality of Gaussian distributions of the movement position of the target object at one time step in the second target time period in the target game. The predicted movement trajectory refers to a movement trajectory determined by the predicted movement position.

[0066] The S203-S204 can predict the moving position of the target object in the target game at each time step in the future time period, thereby accurately representing the moving trajectory of the target object in the target game in the future time period, and improving the prediction effect of the moving trajectory in the target game.

[0067] As an example of S203-S204, on the basis of the above-mentioned example of S201-S202, for each y_t2, t2=21, …, 2d, decompose y_t2 into a predicted mean vector mean_t2, a predicted standard deviation vector scale_t2, and a predicted probability vector logit_t2; determine the predicted moving trajectory y'_T2=[y'_21, …, y'_2d] of the target player in the shooting game within T2 through mean_t2, scale_t2, and logit_t2.

[0068] Referring to Figure 3 , Figure 3 A schematic diagram of a prediction moving trajectory process provided by an embodiment of the present application. The prediction moving trajectory process refers to: first, the object state information, game environment information, and game task information associated with the target object in the target game in the past time period are characterized as object state features, game environment features, and game task features. Second, the object state features, game environment features, and game task features are mixed and encoded into a predicted output vector of the target object in the target game in the future time period. Then, for each predicted output vector, the predicted output vector is decomposed into a predicted mean vector, a predicted standard deviation vector, and a predicted probability vector. Finally, the predicted moving trajectory of the target object in the target game in the future time period is determined through the predicted mean vector, the predicted standard deviation vector, and the predicted probability vector.

[0069] As can be seen from the above technical solutions, by converting the object state information, game environment information, and game task information associated with each time step of the target object in the target game in the past time period into features suitable for the model, and capturing the association relationship between the object state information, game environment information, and game task information in a mixed encoding manner, the moving features of the target object in the target game at each time step in the future time period, i.e., the predicted output vector, are predicted; by decomposing the predicted output vector into a predicted mean vector, a predicted standard deviation vector, and a predicted probability vector, the moving position of the target object in the target game at each time step in the future time period can be predicted, thereby representing the moving trajectory of the target object in the target game in the future time period. Based on this, even if the moving trajectory of the player in the game is very complex, the method can accurately predict the moving trajectory of the player in the game in the future time period, thereby improving the prediction effect of the moving trajectory in the game.

[0070] In the embodiment of the present application, when the predicted moving track of the target object in the second target time period in the target game is determined by predicting the mean vector, the standard deviation vector and the probability vector in S204, it is considered that the moving track of the target player in the target game is very complex, and multiple Gaussian distributions can capture the complexity of the track. The multiple Gaussian distributions are represented by the predicted mean vector and the predicted standard deviation vector to obtain multiple candidate Gaussian distributions to represent the multimodal probability distribution of the moving position of the target object at each time step in the second target time period in the target game. Based on the predicted probability vector and the multiple candidate Gaussian distributions, the predicted moving position of the target object in the second target time period in the target game is sampled to represent the moving position of the target object at each time step in the second target time period in the target game. The predicted moving track of the target object in the second target time period in the target game is determined by the predicted moving position of the target object in the second target time period in the target game. Based on this, the present application provides a possible implementation manner, and S204 includes S2041-S2043 (not shown in the figure).

[0071] S2041: multiple Gaussian distributions are represented according to the predicted mean vector and the predicted standard deviation vector to obtain multiple candidate Gaussian distributions.

[0072] S2042: the multiple candidate Gaussian distributions are sampled and predicted according to the predicted probability vector to obtain the predicted moving position of the target object in the second target time period in the target game.

[0073] S2043: the predicted moving track is determined according to the predicted moving position.

[0074] The multiple candidate Gaussian distributions refer to multiple Gaussian distributions of the moving position of the target object at one time step in the second target time period in the target game. One predicted moving position corresponds to the moving position at one time step in the second target time period.

[0075] The S2041-S2043 can predict the multimodal probability distribution of the moving position of the target object at each time step in the future time period in the target game by implementing multiple Gaussian distribution representation through the decomposed predicted mean vector and the predicted standard deviation vector, and sample and predict the moving position of the target object at each time step in the future time period in the target game based on the multimodal probability distribution and the decomposed predicted probability vector, so as to accurately represent the moving track of the target object in the future time period and improve the prediction effect of the moving track in the game.

[0076] As an example of S2041-S2043, based on the example of S203-S204 described above, the mean_t2 and scale_t2 Gaussian distribution is expressed as a plurality of candidate Gaussian distributions, i.e. N(y_t2|μ_1,Σ_1), …, N(y_t2|μ_I,Σ_I), I is a positive integer, i≥2; based on logit_t2 and N(y_t2|μ_1,Σ_1), …, N(y_t2|μ_I,Σ_I), the predicted movement position of the target player in the shooting game within T2 is sampled as y'_21, …, y'_2d, i.e. the predicted movement trajectory of the target player in the shooting game within T2 y'_T2=[y'_21, …, y'_2d] can be determined.

[0077] In the embodiment of the present application, when the plurality of Gaussian distributions represented by the predicted mean vector and the predicted standard deviation vector are obtained in S2041 described above, it is considered that the Gaussian distribution is determined by the mean and the covariance matrix, and each Gaussian distribution in the plurality of Gaussian distributions has a mean and a covariance matrix. Therefore, the plurality of predicted means are determined by the predicted mean vector, and the plurality of predicted covariance matrices are determined by the predicted standard deviation vector. Then, the plurality of candidate Gaussian distributions are obtained by representing the plurality of predicted means and the plurality of predicted covariance matrices as a plurality of Gaussian distributions. Based on this, the present application provides a possible implementation manner, and S2041 includes S2041a-S2041c (not shown in the figure) as follows.

[0078] S2041a: determining a plurality of predicted means according to the predicted mean vector.

[0079] S2041b: determining a plurality of predicted covariance matrices according to the predicted standard deviation vector.

[0080] S2041c: representing the plurality of predicted means and the plurality of predicted covariance matrices as a plurality of Gaussian distributions to obtain a plurality of candidate Gaussian distributions.

[0081] Wherein, a predicted mean refers to a mean corresponding to a Gaussian distribution of the movement position of the target object in the target game within a time step of the second target time period; and a predicted covariance matrix refers to a covariance matrix corresponding to a Gaussian distribution of the movement position of the target object in the target game within a time step of the second target time period.

[0082] The S2041a-S2041c can accurately realize the plurality of Gaussian distributions by determining the plurality of predicted means corresponding to the predicted mean vector and the plurality of predicted covariance matrices corresponding to the predicted standard deviation vector, so as to accurately predict the multimodal probability distribution of the movement position of the target object in the target game within each time step of the future time period.

[0083] As an example of S2041a-S2041c, based on the example of S2041-S2043 described above, the computer device determines, according to mean_t2, a plurality of predicted mean values u_1, …, u_I, i.e., mean_t2 = [u_1, …, u_I]; determines, according to scale_t2, a plurality of predicted covariance matrices Σ_1, …, Σ_I, i.e., diag(scale_t2^2) = [Σ_1, …, Σ_I], diag() being a diagonal function; and performs a plurality of Gaussian distribution representations on u_1, …, u_I and Σ_1, …, Σ_I to obtain N(y_t2|μ_1, Σ_1), …, N(y_t2|μ_I, Σ_I).

[0084] In the embodiment of the present application, when S2042 above samples the predicted moving position of the target object in the target game in the second target time period based on the target prediction probability vector and the plurality of candidate Gaussian distributions, it is considered that each candidate Gaussian distribution in the plurality of candidate Gaussian distributions has a weight, and the greater the weight, the more likely the candidate Gaussian distribution corresponding to the weight is the probability distribution of the moving position of the target object at each time step in the second target time period in the target game; therefore, the prediction probability vector needs to be weighted first to obtain the prediction weight of each candidate Gaussian distribution, so as to determine the target Gaussian distribution from the plurality of candidate Gaussian distributions through the plurality of prediction weights corresponding to the plurality of candidate Gaussian distributions; and then sample the target Gaussian distribution to obtain the predicted moving position of the target object in the target game in the second target time period. Based on this, the present application provides a possible implementation manner, and S2042 includes the following S2042a-S2042c (not shown in the figure).

[0085] S2042a: weight-distribute the prediction probability vector to obtain the prediction weight of each candidate Gaussian distribution.

[0086] S2042b: determine the target Gaussian distribution from the plurality of candidate Gaussian distributions according to the plurality of prediction weights corresponding to the plurality of candidate Gaussian distributions.

[0087] S2042c: sample and predict the target Gaussian distribution to obtain the predicted moving position.

[0088] wherein one prediction weight refers to the weight corresponding to one Gaussian distribution of the moving position of the target object at one time step in the second target time period in the target game; and the target Gaussian distribution can be the candidate Gaussian distribution corresponding to the maximum prediction weight in the plurality of prediction weights.

[0089] The S2042a-S2042c determines a plurality of prediction weights corresponding to a plurality of candidate Gaussian distributions through a prediction probability vector, can accurately determine a target Gaussian distribution from the plurality of candidate Gaussian distributions to represent a probability distribution of a moving position of a target object in a target game at each time step in a future time period, and thus sample the target Gaussian distribution to predict a predicted moving position of the target object in the target game in a second target time period.

[0090] As an example of the S2042a-S2042c, on the basis of the above-described example of the S2041-S2043, the computer device performs weight distribution on the logit_t2 to obtain a prediction weight of N(y_t2|μ_1,Σ_1) as π_1 and a prediction weight of N(y_t2|μ_I,Σ_I) as π_I, that is, Categorical(logits_t2)=[π_1, …, π_I], Categorical() being a distribution function; and determines a target Gaussian distribution as N(y_t2|μ_i,Σ_i) from N(y_t2|μ_1,Σ_1), …, N(y_t2|μ_I,Σ_I) according to the π_1 of N(y_t2|μ_1,Σ_1), …, the π_I of N(y_t2|μ_I,Σ_I), i∈1, …, I; and performs sampling prediction on N(y_t2|μ_i,Σ_i) to obtain y'_t2, that is, y'_t2=sample(π_i×N(y_t2|μ_i,Σ_i)), sample() being a sampling function.

[0091] In the embodiment of the present application, when the object state feature, the game environment feature, and the game task feature are mixed and encoded as the predicted output vector of the target object in the target game in the second target time period in the above-described S202, considering the association relationship among the object state information, the game environment information, and the game task information, the object state feature, the game environment feature, and the game task feature can be spliced into a target splicing vector first, so that the target splicing vector includes the object state feature, the game environment feature, and the game task feature; and then the target splicing vector is mixed and encoded to capture the association relationship among the object state information, the game environment information, and the game task information, thereby predicting the predicted output vector of the target object in the target game in the second target time period. Based on this, the present application provides a possible implementation manner, and the S202 includes the following S2021-S2022 (not shown in the figure).

[0092] S2021: performing feature splicing on the object state feature, the game environment feature, and the game task feature to obtain a target splicing vector.

[0093] S2022: performing mixed encoding on the target splicing vector to obtain a predicted output vector.

[0094] wherein one target splicing vector comprises the features representing the object state information, the features representing the game environment information and the features representing the game task information corresponding to one time step in the first target time period.

[0095] The S2021-S2022 splices the object state features, the game environment features and the game task features into target splicing vectors, realizes the association of the object state information, the game environment information and the game task information, so as to accurately capture the association relationship between the object state information, the game environment information and the game task information in a hybrid encoding manner, and thus accurately predict the moving features of the target object in each time step in the future time period in the target game.

[0096] As an example of S2021-S2022, on the basis of the above-mentioned example of S202, the computer device splices o_11, …, o_1c, g_11, …, g_1c and m_11, …, m_1c to obtain c target splicing vectors x_11, …, x_1c, wherein x_11 comprises o_11, g_11 and m_11, and x_1c comprises o_1c, g_1c and m_1c; and the computer device hybrid encodes x_11, …, x_1c to obtain y_21, …, y_2d.

[0097] In the above-mentioned S2022, when the target splicing vector is hybrid encoded into the predicted output vector of the target object in the second target time period in the target game, at least the following implementation manners are adopted.

[0098] One implementation manner is that the self-attention mechanism is used to capture the dependency relationship between the input representations in the input sequence; therefore, for the target splicing vector, the self-attention mechanism hybrid encoding can be used to capture the association relationship between the object state information, the game environment information and the game task information, so as to predict the predicted output vector of the target object in the second target time period in the target game. Based on this, the present application provides a possible implementation manner, S2022 comprises S2022a (not shown in the figure): hybrid encoding of the target splicing vector based on self-attention to obtain the predicted output vector.

[0099] The S2022a hybrid encodes the target splicing vector through the self-attention mechanism, captures the dependency relationship between the object state information, the game environment information and the game task information, captures the association relationship between the object state information, the game environment information and the game task information, and thus more accurately predicts the moving features of the target object in each time step in the future time period in the target game.

[0100] As an example of S2022a, based on the above examples of S2021-S2022, the computer device performs self-attention-based hybrid encoding on x_11, …, x_1c to obtain y_21, …, y_2d. The specific steps are: for each target splicing vector x_t1, t1 = 11, …, 1c, map x_t1 into three vectors of Query, Key and Value, denoted as Q, K and V respectively, Q = W_q x x_t1 + b_q, K = W_k x x_t1 + b_k, V = W_v x x_t1 + b_v, W_q, W_k, W_v are weights, and b_q, b_k, b_v are biases; calculate the self-attention feature through Q, K and V, i.e. AttentionHead(Q, K, V) = softmax(Q x K^T / sqrt(d)) x V, d is the dimension of x_11, …, x_1c; convert AttentionHead(Q, K, V) into the target encoding feature z_t2 of each time step in T2 through a feedforward neural network, i.e. FNN(AttentionHead(Q, K, V)); convert z_t2 into y_t2 through a linear layer, i.e. y_t2 = W x z_t2 + b, W is a weight, and b is a bias.

[0101] Another implementation manner is: considering that the multi-head attention mechanism is used to capture complex dependency relationships between input representations in the input sequence; therefore, for the target splicing vector, hybrid encoding can be performed through the multi-head attention mechanism to capture complex association relationships between the object state information, the game environment information and the game task information, so as to predict the predicted output vector of the target object in the target game in the second target time period. Based on this, the present application provides a possible implementation manner, S2022 includes S2022b (not shown in the figure): performing multi-head attention-based hybrid encoding on the target splicing vector to obtain the predicted output vector.

[0102] The S2022b hybrid encodes the target splicing vector through the multi-head attention mechanism to capture complex dependency relationships between the understanding of the object state information, the game environment information and the game task information, so as to capture complex association relationships between the object state information, the game environment information and the game task information, thereby further accurately predicting the movement feature of the target object in the target game at each time step in the future time period.

[0103] As an example of S2022b, based on the above examples of S2021-S2022, the computer device performs multi-head attention-based mixed encoding on x_11, …, x_1c to obtain y_21, …, y_2d. The specific steps are as follows: for each target splicing vector x_t1, t1 = 11, …, 1c, map x_t1 into multiple groups of query (Q), key (K) and value (V) vectors, denoted as multiple groups of Q, K and V; calculate the multi-head attention feature through multiple groups of Q, K and V, that is, MultiHead(Q, K, V) = Concat(head_1, …, head_h) × W^O, head_1, …, head_h are the attention weights corresponding to h heads respectively, and W^O is the output weight matrix; convert MultiHead(Q, K, V) into the target encoding feature z_t2 of each time step in T2 through a feedforward neural network, that is, FNN(MultiHead(Q, K, V)); convert z_t2 into y_t2 through a linear layer, that is, y_t2 = W × z_t2 + b, W is a weight, and b is a bias.

[0104] In the embodiment of the present application, when the above S2021 splicing object state features, game environment features and game task features are implemented as target splicing vectors, in order to enable subsequent perception and capture of the position information of each element in the target splicing vector, the object state features, game environment features and game task features need to be spliced into a spliced vector first, so that the spliced vector includes the object state features, game environment features and game task features, and a position encoding vector is obtained by position encoding of the spliced vector; then, the spliced vector and the position encoding vector are spliced, and the target splicing vector can be obtained. Based on this, the present application provides a possible implementation manner, and S2021 includes the following S2021a-S2021c (not shown in the figure).

[0105] S2021a: Splice the object state features, game environment features and game task features to obtain a spliced vector.

[0106] S2021b: Position encoding of the spliced vector to obtain a position encoding vector.

[0107] S2021c: Vector splicing of the spliced vector and the position encoding vector to obtain a target splicing vector.

[0108] Among them, one spliced vector includes features corresponding to one time step in the first target time period, representing object state information, game environment information and game task information; and one position encoding vector represents the position information of each feature in one spliced vector.

[0109] The S2021a-S2021c obtains the position encoding vector corresponding to the spliced vector based on the spliced object state feature, the spliced game environment feature and the spliced game task feature, and can represent the position information of each feature in the spliced vector; and splices the spliced vector and the position encoding vector, so that the target spliced vector not only associates the object state information, the game environment information and the game task information in the feature, but also represents the position information, so as to more accurately capture the association relationship between the object state information, the game environment information and the game task information in the mixed encoding manner.

[0110] As an example of S2021a-S2021c, based on the example of S2021-S2022, the computer device splices o_11, …, o_1c, g_11, …, g_1c and m_11, …, m_1c to obtain a plurality of spliced vectors x'_11, …, x'_1c, i.e., x'_11=[o_11, g_11, m_11], …, x'_1c=[o_1c, g_1c, m_1c]; for each spliced vector x'_t1, t1=11, …, 1c, the position encoding is performed on x'_t1 to obtain the position encoding vector P_t1 corresponding to x'_t1, i.e., PE_t1=[sin(t1 / 10000^(2j / d)), cos(t1 / 10000^(2j / d))], d is the dimension of x'_t1; and the vector splicing is performed on x'_11, …, x'_1c and PE_11 corresponding to x'_11, …, PE_1c corresponding to x'_1c to obtain a plurality of target spliced vectors x_11, …, x_1c, i.e., x_11=x'_11+PE_11, …, x_1c=x'_1c+PE_1c.

[0111] In the embodiment of the present application, when the target game is a game, not only the first player state information of the target player in the past time period in the target game affects the moving track of the target player in the future time period, but also the second player state information of the game player affects the moving track of the target player in the future time period, i.e., the object state information associated with the target object in the first target time period in the target game needs to include the first object state information of the target object and the second object state information of the game object in the first target time period in the target game. Based on this, the present application provides a possible implementation manner, the target game is a game, and the object state information includes the first object state information of the target object and the second object state information of the game object.

[0112] In the embodiments of the present application, the target mixture coding model and the target Gaussian mixture model in S201-S206 are obtained by training an initial mixture coding model and an initial Gaussian mixture model according to a plurality of preset object state information associated with a preset object in a first historical time period in a preset game, a plurality of preset game environment information and a plurality of preset game task information, and an actual moving track of the preset object in a second historical time period in the preset game, the second historical time period being after the first historical time period.

[0113] The training process refers to: first, the sample object state information, the sample game environment information and the sample game task information associated with a sample object in a first historical time period in a sample game need to be expressed as sample object state features, sample game environment features and sample game task features that can be processed by the model, wherein the sample object state information includes sample moving state information; and the sample object state features, the sample game environment features and the sample game task features are input into the initial coding model for mixture coding to capture the association between the sample object state information, the sample game environment information and the sample game task information, so as to predict a sample prediction output vector of the sample object in a second historical time period in the sample game to represent the moving features of the sample object at each time step in the second historical time period in the sample game.

[0114] Then, for each sample prediction output vector, the sample prediction output vector is decomposed into parameters of the initial Gaussian model, i.e., a sample prediction mean vector, a sample prediction standard deviation vector and a sample prediction probability vector; the prediction probability density of the moving position of the sample object in the second historical time period in the sample game can be predicted by the sample prediction mean vector, the sample prediction standard deviation vector and the sample prediction probability vector, to represent the prediction probability distribution of the moving position of the sample object at each time step in the second historical time period in the sample game.

[0115] Finally, the initial coding model and the initial Gaussian model are trained by the prediction probability density and the actual moving track of the sample object in the second historical time period in the sample game, so that the prediction probability density conforms to the actual moving track, until the model converges or the training number reaches the completion of the training, and the trained initial coding model and the trained initial Gaussian model are used as the mixture coding model and the Gaussian mixture model for moving track prediction. Based on this, a possible implementation manner is provided, and the method further includes S1-S5 (not shown in the figure).

[0116] S1: The sample object state information, the sample game environment information and the sample game task information associated with a sample object in a first historical time period in a sample game are feature-expressed to obtain sample object state features, sample game environment features and sample game task features; the sample object state information includes sample moving state information.

[0117] S2: encoding the sample object state feature, the sample game environment feature and the sample game task feature through the initial encoding model to obtain a sample prediction output vector of the sample object in a second historical time period in the sample game; the second historical time period is after the first historical time period.

[0118] Wherein, the sample game is a game that a game player needs to move; the first historical time period refers to a past time period; the sample object refers to a game player; one sample object state information corresponds to one time step in the first historical time period, which refers to movement state information and life state information, etc., the movement state information refers to one or more of position information, speed information, acceleration information and orientation information, and the life state information refers to one or more of blood volume information and life value information; one sample game environment information corresponds to one time step in the first historical time period; one sample game task information corresponds to one time step in the first historical time period; one sample object state feature refers to a feature representing the sample object state information corresponding to one time step in the first historical time period; one sample game environment feature refers to a feature representing the sample game environment information corresponding to one time step in the first historical time period; one sample game task feature refers to a feature representing the sample game task information corresponding to one time step in the first historical time period; the second historical time period refers to another past time period; the sample prediction output vector corresponds to one time step in the second historical time period, which refers to a movement feature.

[0119] S3: for each sample prediction output vector, performing vector decomposition on the sample prediction output vector to obtain a sample prediction mean vector, a sample prediction standard deviation vector and a sample prediction probability vector of the initial Gaussian model.

[0120] S4: obtaining a prediction probability density of the moving position of the sample object in the second historical time period in the sample game according to the sample prediction mean vector, the sample prediction standard deviation vector and the sample prediction probability vector.

[0121] Wherein, one sample prediction mean vector includes a plurality of mean vectors corresponding to a plurality of Gaussian distributions of the moving position of the sample object in one time step in the second historical time period in the sample game; one sample prediction standard deviation vector includes a plurality of standard deviation vectors corresponding to a plurality of Gaussian distributions of the moving position of the sample object in one time step in the second historical time period in the sample game; one sample prediction probability vector includes a plurality of weights corresponding to a plurality of Gaussian distributions of the moving position of the sample object in one time step in the second historical time period in the sample game; the prediction probability density refers to the probability density of the predicted moving position.

[0122] S5: training the initial encoding model and the initial Gaussian model according to the predicted probability density and the actual moving trajectory of the sample object in the second historical time period in the sample game, to obtain a hybrid encoding model and a Gaussian mixture model for moving trajectory prediction.

[0123] The actual moving trajectory refers to a moving trajectory determined by a plurality of actual moving positions corresponding to a plurality of time steps.

[0124] The S1-S5 converts the sample object state information, sample game environment information and sample game task information of each time step associated with the sample object in the past time period in the sample game into features suitable for the model, so that the model can process the sample object state information, sample game environment information and sample game task information; on this basis, the association between the sample object state information, sample game environment information and sample game task information is captured in a hybrid encoding manner, so as to predict the moving feature of the sample object in each time step in the past time period in the sample game, i.e. the sample prediction output vector, to provide a data basis for subsequent prediction of the probability distribution of the moving position of the sample object in each time step in the past time period in the sample game. Even if the moving trajectory of the sample player in the sample game is very complex, by decomposing the sample prediction output vector into the sample prediction mean vector, sample prediction standard deviation vector and sample prediction probability vector of the initial Gaussian model, the probability density of the moving position of the sample object in each time step in the past time period in the sample game is predicted. The training direction of the predicted probability density conforms to the actual moving trajectory of the sample object in the past time period in the sample game, and the initial encoding model and the initial Gaussian model are trained, so that the hybrid encoding model and the Gaussian mixture model which can accurately predict the moving trajectory of the player in the future time period even if the moving trajectory of the player in the game is very complex can be accurately trained.

[0125] As an example of S1-S5, the sample game is a shooting game, the first historical time period is T3, T3 includes c time steps, c is a positive integer, c≥2, the sample object is a sample player, the initial encoding model is an initial Transformer model, the second historical time period is T4, T4 is after T3, T4 includes d time steps, d is a positive integer, d≥2, the initial Gaussian model represents I Gaussian models, I is a positive integer, I≥2, and the actual moving trajectory is y_T4=[y_41,…,y_4d].

[0126] The computer device performs feature representation on the c sample object state information, the c sample game environment information and the c sample game task information associated with the T1-in sample player in the shooting game, obtains c sample object state features o_31, …, o_3c, c sample game environment features g_31, …, g_3c, and c sample game task features m_31, …, m_3c; the sample object state information includes sample movement state information; through the initial Transformer model, the o_31, …, o_3c, g_31, …, g_3c and m_31, …, m_3c are mixedly coded, and d sample prediction output vectors y'_41, …, y'_4d of the T4-in sample player in the shooting game are obtained.

[0127] For each y'_t4, t4=41, …, 4d, the vector decomposition is performed on y'_t4, and the sample prediction mean vector of the initial Gaussian model is obtained as mean_t4, the sample prediction standard deviation vector is scale_t4, and the sample prediction probability vector is logit_t4; according to mean_t4, scale_t4 and logit_t4, the prediction probability density of the movement position of the T4-in sample player in the shooting game is obtained as p(y'_t4).

[0128] According to p(y'_t4) and y_T4=[y_41, …, y_4d] of the T4-in sample player in the shooting game, the initial Transformer model and the initial Gaussian mixture model are trained, and a mixed coding model is obtained as a target Transformer model and a Gaussian mixture model.

[0129] Referring to Figure 4 , Figure 4A schematic diagram of a process for training a mixed encoding model and a Gaussian mixed model is provided in the embodiments of the present application. The process for training a mixed encoding model and a Gaussian mixed model refers to the following. First, sample object state information, sample game environment information and sample game task information associated with a sample object in a past time period in a sample game are expressed as sample object state features, sample game environment features and sample game task features. Second, the sample object state features, sample game environment features and sample game task features are input into an initial encoding model to encode a sample prediction output vector of the sample object in a subsequent past time period in the sample game. Then, for each sample prediction output vector, the sample prediction output vector is decomposed into a sample prediction mean vector, a sample prediction standard deviation vector and a sample prediction probability vector of an initial Gaussian model; the prediction probability density of the moving position of the sample object in a second historical time period in the sample game is predicted based on the sample prediction mean vector, the sample prediction standard deviation vector and the sample prediction probability vector. Finally, the initial encoding model and the initial Gaussian model are trained based on a plurality of prediction probability densities and actual moving trajectories of the sample object in the subsequent past time period in the sample game, and the trained initial mixed encoding model and the trained initial Gaussian mixed model are used as the mixed encoding model and the Gaussian mixed model.

[0130] In the embodiments of the present application, when the prediction probability density of the moving position of the sample object in the second historical time period in the sample game is implemented based on the sample prediction mean vector, the sample prediction standard deviation vector and the sample prediction probability vector in S4, the moving trajectory of the sample player in the sample game is very complex, and a plurality of Gaussian distributions can capture the complexity of the trajectory. Therefore, the sample prediction mean vector and the sample prediction standard deviation vector are input into the initial Gaussian model to represent a plurality of Gaussian distributions, and a plurality of sample Gaussian distributions are output to represent the multimodal probability distribution of the moving position of the sample object at each time step in the second historical time period in the sample game. The sample prediction probability vector is weighted to obtain the prediction weight of each sample Gaussian distribution, so that the prediction probability density of the moving position of the sample object in the second historical time period in the sample game is calculated by weighting the plurality of sample Gaussian distributions and the plurality of prediction weights corresponding to the plurality of sample Gaussian distributions, to represent the prediction probability distribution of the moving position of the sample object at each time step in the second historical time period. Based on this, the present application provides a possible implementation manner, and S4 includes the following S41-S43 (not shown in the figure).

[0131] S41: A plurality of sample Gaussian distributions are obtained by representing a plurality of Gaussian distributions based on the sample prediction mean vector and the sample prediction standard deviation vector through the initial Gaussian model.

[0132] S42: The sample prediction probability vector is weighted to obtain the sample weight of each sample Gaussian distribution.

[0133] S43: Obtain a prediction probability density by weighted calculation according to the plurality of sample Gaussian distributions and the plurality of sample Gaussian distribution corresponding plurality of Ying weights.

[0134] Wherein, the plurality of sample Gaussian distributions refers to a plurality of Gaussian distributions of the moving position of the sample object in the sample game at a time step in the second historical time period; one sample weight refers to a weight corresponding to one Gaussian distribution of the moving position of the sample object in the sample game at a time step in the second historical time period.

[0135] The S41-S43 can realize the plurality of Gaussian distribution representation through the decomposed sample prediction mean vector and sample prediction standard deviation vector, can predict the probability distribution of the moving position of the preset object in the sample game at each time step in the past time period, and can combine the plurality of prediction weights corresponding to the decomposed sample prediction probability vector to perform weighted calculation to predict the probability density of the moving position of the sample object in the sample game at each time step in the past time period.

[0136] As an example of S41-S43, on the basis of the above-mentioned example of S1-S5, the plurality of Gaussian distribution representation is obtained according to mean_t4 and scale_t4 through the initial Gaussian model, I candidate Gaussian distributions N(y'_t4|μ_1,Σ_1), …, N(y'_t4|μ_I,Σ_I) are obtained; the weight distribution of logit_t4 is obtained, the prediction weight of N(y'_t4|μ_1,Σ_1) is π_1, the prediction weight of N(y'_t4|μ_I,Σ_I) is π_I, that is, Categorical(logits_t4)=[π_1,…,π_I], Categorical() is a distribution function; weighted calculation is performed according to (y'_t4|μ_1,Σ_1), …, N(y'_t4|μ_I,Σ_I) and π_1 of N(y'_t4|μ_1,Σ_1), …, π_I of N(y'_t4|μ_I,Σ_I), to obtain p(y'_t4)=Σ(π_i×N(y'_t4|μ_i,Σ_i)).

[0137] In the embodiments of the present application, when the mixed encoding model and the Gaussian mixture model are specifically implemented by training the initial encoding model and the initial Gaussian model based on the predicted probability density and the actual moving track of the sample object in the sample game in the second historical time period, considering that the training direction is to make the predicted probability density consistent with the actual moving track, the predicted probability of the predicted probability density on the actual moving track needs to be determined first, and then the initial encoding model and the initial Gaussian model are trained by maximizing each predicted probability, so that the predicted probability density is consistent with the actual moving track, until the model converges or the training number reaches the completion of training, so as to obtain the mixed encoding model and the Gaussian mixture model by using the trained initial encoding model and the trained initial Gaussian model. Based on this, the present application provides a possible implementation manner, and S5 includes S51-S52 (not shown in the figure).

[0138] S51: determining the predicted probability of the predicted probability density on the actual moving track.

[0139] S52: training the initial encoding model and the initial Gaussian model by maximizing each predicted probability to obtain the mixed encoding model and the Gaussian mixture model.

[0140] In the embodiments of the present application, when the mixed encoding model and the Gaussian mixture model are specifically implemented by training the initial encoding model and the initial Gaussian model based on the predicted probability density and the actual moving track of the sample object in the sample game in the second historical time period, considering that the training direction is to make the predicted probability density consistent with the actual moving track, the predicted probability of the predicted probability density on the actual moving track needs to be determined first, and then the initial encoding model and the initial Gaussian model are trained by maximizing each predicted probability, so that the predicted probability density is consistent with the actual moving track, until the model converges or the training number reaches the completion of training, so as to obtain the mixed encoding model and the Gaussian mixture model by using the trained initial encoding model and the trained initial Gaussian model. Based on this, the present application provides a possible implementation manner, and S5 includes S51-S52 (not shown in the figure).

[0141] The S51-S52 determines the predicted probability of the predicted probability density on the actual moving track, which can predict the probability that the moving position of the sample object at each time step in the past time period after the sample game is the actual moving position of the sample object at the time step in the actual moving track. By maximizing each predicted probability, the moving position of the sample object at each time step in the past time period after the sample game is the actual moving position of the sample object at the time step in the actual moving track, and the initial encoding model and the initial Gaussian model are trained, which can more accurately train the mixed encoding model and the Gaussian mixture model that can accurately predict the moving track of the player in the future time period in the game even if the moving track of the player in the game is very complex.

[0142] As an example of S51-S52, based on the example of S1-S5, the computer device determines d predicted probabilities of p(y'_t4) on y_T4 as p(y'_41|y_41), … p(y'_4d|y_4d). By maximizing each predicted probability p(y'_t4|y_t4), i.e. minimizing ﹣log(p(y'_t4)), the initial Transformer model and the initial Gaussian model are trained to obtain the target Transformer model and the Gaussian mixture model.

[0143] Further, in the embodiments of the present application, after the predicted moving track of the target object in the second target time period in the target game is determined in S204, in order to realize the downstream application of the moving track correlation through the predicted moving track of the target object in the second target time period in the target game, the game object can also be subjected to the downstream application based on the moving track through the predicted moving track, and the downstream application result is obtained. Based on this, the present application provides a possible implementation manner, and the method further comprises S6 (not shown in the figure): the target object is subjected to the downstream application based on the moving track according to the predicted moving track, and the downstream application result is obtained.

[0144] Referring to Figure 5 , Figure 5 A schematic diagram of the downstream application of the moving track is provided for the embodiments of the present application. After the predicted moving track of the target player in the future time period in the target game is determined through the player state information, the game environment information and the game task information associated with the target player in the past time period in the target game, the predicted moving track can be applied to the related game artificial intelligence tasks such as the game artificial intelligence training task based on supervised learning, the game artificial intelligence training task based on reinforcement learning, the player data analysis task and the game anti-cheating task.

[0145] For example, in the game artificial intelligence training task, the game artificial intelligence can be made to move along the route as close as possible to the predicted moving track of the player through the predicted moving track of the player, so as to optimize the route humanization result of the game artificial intelligence and improve the application range and application scenario of the game artificial intelligence. In the player data analysis task, the predicted moving track of most players can be used to well judge whether the game design meets the expectation, and then the game design is improved. In the game anti-cheating task, the predicted moving track of the player can be used to judge whether the moving track exceeds the expectation, so as to judge whether the player has the cheating behavior.

[0146] It should be noted that the implementation manners provided by the present application in the above aspects can be further combined to provide more implementation manners.

[0147] Based on Figure 2 Corresponding to the moving track prediction method provided by the embodiments, the present application further provides a moving track prediction device, referring to Figure 6 , Figure 6 A structure diagram of the moving track prediction device provided by the embodiments of the present application is shown in the figure, and the moving track prediction device 600 comprises a feature representation unit 601, a mixed coding unit 602, a vector decomposition unit 603, a Gaussian distribution representation unit 604, a sampling prediction unit 605 and a determination unit 606.

[0148] The feature representation unit 601 is configured to perform feature representation on object state information, game environment information, and game task information associated with the target object in a first target time period of the target game, to obtain object state features, game environment features, and game task features. The object state information includes movement trajectory information.

[0149] The mixed encoding unit 602 is configured to perform mixed encoding on the object state features, the game environment features, and the game task features, to obtain a prediction output vector of the target object in a second target time period of the target game. The second target time period is after the first target time period.

[0150] The vector decomposition unit 603 is configured to perform vector decomposition on each prediction output vector to obtain a prediction mean vector, a prediction standard deviation vector, and a prediction probability vector.

[0151] The determination unit 604 is configured to determine a prediction movement trajectory of the target object in the second target time period of the target game according to the prediction mean vector, the prediction standard deviation vector, and the prediction probability vector.

[0152] In a possible implementation, the determination unit 604 is configured to:

[0153] perform a plurality of Gaussian distribution representations according to the prediction mean vector and the prediction standard deviation vector, to obtain a plurality of candidate Gaussian distributions;

[0154] perform sampling prediction on the plurality of candidate Gaussian distributions according to the prediction probability vector, to obtain a prediction movement position of the target object in the second target time period of the target game;

[0155] determine the prediction movement trajectory according to the prediction movement position.

[0156] In a possible implementation, the determination unit 604 is configured to:

[0157] determine a plurality of prediction means according to the prediction mean vector;

[0158] determine a plurality of prediction covariance matrices according to the prediction standard deviation vector;

[0159] perform a plurality of Gaussian distribution representations on the plurality of prediction means and the plurality of prediction covariance matrices, to obtain a plurality of candidate Gaussian distributions.

[0160] In a possible implementation, the determination unit 604 is configured to:

[0161] perform weight distribution on the prediction probability vector, to obtain a prediction weight of each candidate Gaussian distribution;

[0162] determine a target Gaussian distribution from the plurality of candidate Gaussian distributions according to a plurality of prediction weights corresponding to the plurality of candidate Gaussian distributions;

[0163] Sample the target Gaussian distribution to obtain a predicted moving position.

[0164] In a possible implementation, the hybrid encoding unit 602 is configured to:

[0165] concatenate the object state feature, the game environment feature and the game task feature to obtain a target concatenated vector;

[0166] perform hybrid encoding on the target concatenated vector to obtain a predicted output vector.

[0167] In a possible implementation, the hybrid encoding unit 602 is configured to:

[0168] perform self-attention-based hybrid encoding on the target concatenated vector to obtain the predicted output vector.

[0169] In a possible implementation, the hybrid encoding unit 602 is configured to:

[0170] perform multi-head attention-based hybrid encoding on the target concatenated vector to obtain the predicted output vector.

[0171] In a possible implementation, the hybrid encoding unit 602 is configured to:

[0172] concatenate the object state feature, the game environment feature and the game task feature to obtain a concatenated vector;

[0173] perform position encoding on the concatenated vector to obtain a position encoding vector;

[0174] concatenate the concatenated vector and the position encoding vector to obtain the target concatenated vector.

[0175] In a possible implementation, the target game is a turn-based game, and the object state information includes first object state information of a target object and second object state information of a turn-based object.

[0176] In a possible implementation, the apparatus further includes a training unit.

[0177] The training unit is configured to:

[0178] perform feature representation on sample object state information, sample game environment information and sample game task information associated with a sample object in a first historical time period in a sample game to obtain sample object state feature, sample game environment feature and sample game task feature; the sample object state information includes sample moving state information.

[0179] The initial encoding model is used to encode the sample object state features, the sample game environment features, and the sample game task features to obtain a sample prediction output vector of the sample object in a second historical time period in the sample game, the second historical time period being after the first historical time period;

[0180] For each sample prediction output vector, the sample prediction output vector is decomposed to obtain a sample prediction mean vector, a sample prediction standard deviation vector, and a sample prediction probability vector of the initial Gaussian model;

[0181] According to the sample prediction mean vector, the sample prediction standard deviation vector, and the sample prediction probability vector, a prediction probability density of a moving position of the sample object in the second historical time period in the sample game is obtained.

[0182] According to the prediction probability density and an actual moving trajectory of the sample object in the second historical time period in the sample game, the initial encoding model and the initial Gaussian model are trained to obtain a hybrid encoding model and a Gaussian mixture model for moving trajectory prediction.

[0183] In a possible implementation, the training unit is configured to:

[0184] According to the sample prediction mean vector and the sample prediction standard deviation vector, a plurality of Gaussian distributions are represented by the initial Gaussian model to obtain a plurality of sample Gaussian distributions.

[0185] The sample prediction probability vector is weighted to obtain a sample weight of each sample Gaussian distribution.

[0186] According to the plurality of sample Gaussian distributions and the plurality of sample weights corresponding to the plurality of sample Gaussian distributions, weighted calculation is performed to obtain the prediction probability density.

[0187] In a possible implementation, the training unit is configured to:

[0188] The prediction probability density is determined on the actual moving trajectory.

[0189] By maximizing each prediction probability, the initial encoding model and the initial Gaussian model are trained to obtain the hybrid encoding model and the Gaussian mixture model.

[0190] In a possible implementation, the apparatus further includes an analysis unit.

[0191] The analysis unit is configured to perform a moving trajectory-based downstream application on the target object according to the prediction moving trajectory to obtain a downstream application result.

[0192] It can be seen from the technical solution that the mobile trajectory prediction device comprises a feature representation unit, a hybrid coding unit, a vector decomposition unit and a determination unit. The feature representation unit represents object state information, game environment information and game task information associated with a target object in a first target time period in a target game as object state features, game environment features and game task features, wherein the object state information comprises movement state information; the hybrid coding unit hybrid-codes the object state features, the game environment features and the game task features into a prediction output vector of the target object in a second target time period after the first target time period in the target game. By converting the object state information, the game environment information and the game task information associated with each time step of the target object in the past time period in the target game into features suitable for the model, and capturing the association relationship between the object state information, the game environment information and the game task information in a hybrid coding manner, the movement features of the target object in each time step in the future time period in the target game, i.e., the prediction output vector, are predicted.

[0193] The vector decomposition unit decomposes each prediction output vector into a prediction mean vector, a prediction standard deviation vector and a prediction probability vector; and the determination unit determines a prediction mobile trajectory of the target object in the second target time period in the target game based on the prediction mean vector, the prediction standard deviation vector and the prediction probability vector. By decomposing the prediction output vector into the prediction mean vector, the prediction standard deviation vector and the prediction probability vector, the movement position of the target object in each time step in the future time period in the target game can be predicted, thereby representing the movement trajectory of the target object in the future time period. Based on this, even if the movement trajectory of the player in the game is very complex, the device can accurately predict the movement trajectory of the player in the future time period in the game, thereby improving the prediction effect of the movement trajectory in the game.

[0194] The embodiments of the present application also provide a computer device, which can be a server, as shown in Figure 7 , Figure 7 A structural diagram of a server provided by the embodiments of the present application is shown in FIG. 7. The server 700 can have great differences due to different configurations or performances, and can comprise one or more processors, such as a central processing unit CPU 722, and a memory 732, one or more storage media 730 (such as one or more mass storage devices) storing application programs 742 or data 744. The memory 732 and the storage media 730 can be temporary storage or persistent storage. The programs stored in the storage media 730 can comprise one or more modules (not shown in the figure), each of which can comprise a series of instruction operations in the server. Further, the central processing unit 722 can be configured to communicate with the storage media 730 to execute the series of instruction operations in the storage media 730 on the server 700.

[0195] The server 700 can also include one or more power supplies 726, one or more wired or wireless network interfaces 750, one or more input / output interfaces 758, and / or one or more operating systems 741, such as Windows Server TM , Mac OS X TM , Unix TM , Linux TM , FreeBSD TM , and the like.

[0196] In this embodiment, the central processing unit 722 in the server 700 can execute the method provided in the various optional implementations of the above embodiments.

[0197] The computer device provided by the embodiments of the present application can also be a terminal. Referring to Figure 8 , Figure 8 is a structural diagram of a terminal provided by the embodiments of the present application. Taking a smart phone as an example, the smart phone includes a radio frequency (RF) circuit 810, a memory 820, an input unit 830, a display unit 840, a sensor 850, an audio circuit 860, a wireless fidelity (WiFi) module 870, a processor 880, and a power supply 890, and the like. The input unit 830 can include a touch panel 831 and other input devices 832, and the display unit 840 can include a display panel 841. The audio circuit 860 can include a speaker 861 and a microphone 862. Those skilled in the art can understand that the structure of the smart phone shown in Figure 8 does not constitute a limitation on the smart phone, and can include more or fewer components than those shown in the figure, or combine certain components, or different arrangement of components.

[0198] The memory 820 can be used to store software programs and modules, and the processor 880 executes various function applications and data processing of the smart phone by running the software programs and modules stored in the memory 820. The memory 820 can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, application programs required by at least one function (such as a sound playing function, an image playing function, etc.), and the like; and the data storage area can store data created according to the use of the smart phone (such as audio data, a phone book, etc.), and the like. In addition, the memory 820 can include a high-speed random access memory, and can also include a non-volatile memory, such as at least one magnetic disk storage device, a flash memory device, or other volatile solid-state memory device.

[0199] The processor 880 is the control center of the smartphone, connecting all parts of the smartphone through various interfaces and lines, executing various functions of the smartphone and processing data by running or executing software programs and / or modules stored in the memory 820 and calling data stored in the memory 820. Optionally, the processor 880 can include one or more processing units; preferably, the processor 880 can integrate an application processor and a modem processor, wherein the application processor mainly processes the operating system, user interface and application programs, and the modem processor mainly processes wireless communication. It can be understood that the above-mentioned modem processor can also not be integrated into the processor 880.

[0200] In the embodiment, the processor 880 in the smartphone can execute the method provided in various optional implementation manners of the above-mentioned embodiments.

[0201] According to an aspect of the present application, a computer readable storage medium is provided, the computer readable storage medium is used to store a computer program, when the computer program is running on a computer device, the computer program causes the computer device to execute the method provided in various optional implementation manners of the above-mentioned embodiments.

[0202] According to an aspect of the present application, a computer program product is provided, the computer program product includes a computer program stored in a computer readable storage medium. The processor of the computer device reads the computer program from the computer readable storage medium, and the processor executes the computer program, so that the computer device executes the method provided in various optional implementation manners of the above-mentioned embodiments.

[0203] The description of the corresponding flow or structure of each of the above-mentioned figures has its own emphasis, and the part not described in detail in a certain flow or structure can refer to the related description of other flows or structures.

[0204] The terms "first", "second", and the like in the specification of the present application and the above-mentioned drawings are used to distinguish similar objects, and do not necessarily indicate a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not necessarily limit to those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0205] In several embodiments provided in the present application, it should be understood that the disclosed system, device and method can be implemented in other manners. For example, the described device embodiments are merely schematic. For example, the division of the units is merely a logical function division. There can be another division manner for the actual implementation, for example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed mutual couplings or direct couplings or communication connections between different units, can be indirect couplings or communication connections through some interfaces, devices or units, and can be in electrical, mechanical or other forms.

[0206] The units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, i.e., can be located in one place, or can be distributed on multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the embodiment.

[0207] In addition, each functional unit in the various embodiments of the present application can be integrated into a processing unit, or each unit can exist physically, or two or more units can be integrated into one unit. The integrated unit can be implemented in the form of hardware, or in the form of a software functional unit.

[0208] The integrated unit, if implemented in the form of a software functional unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application essentially or the part that contributes to the prior art, or all or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes a number of instructions for causing a computer device to execute all or part of the steps of the methods described in the various embodiments of the present application. The foregoing storage medium includes: U disk, mobile hard disk, read-only memory (ROM), RAM, magnetic disk or optical disk, and various computer program storage media.

[0209] In the embodiments of the present application, the term "module" or "unit" refers to a computer program or a part of a computer program with a predetermined function, and works with other related parts to achieve a predetermined target, and can be implemented entirely or partially by using software, hardware (such as a processing circuit or a memory) or a combination thereof. Similarly, one processor (or multiple processors or memories) can be used to implement one or more modules or units. In addition, each module or unit can be a part of an overall module or unit that includes the functions of the module or unit.

[0210] The above-described embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for predicting movement trajectories, characterized in that, The method includes: Feature representation is performed on the object state information, game environment information, and game task information associated with the target object within the first target time period in the target game to obtain object state features, game environment features, and game task features; the object state information includes movement state information. The object state features, the game environment features, and the game task features are mixed and encoded to obtain the predicted output vector of the target object in the second target time period in the target game; the second target time period is after the first target time period. For each of the predicted output vectors, the predicted output vectors are decomposed to obtain the predicted mean vector, the predicted standard deviation vector, and the predicted probability vector. Based on the predicted mean vector, the predicted standard deviation vector, and the predicted probability vector, the predicted movement trajectory of the target object within the second target time period in the target game is determined.

2. The method according to claim 1, characterized in that, Determining the predicted movement trajectory of the target object within the second target time period in the target game based on the predicted mean vector, the predicted standard deviation vector, and the predicted probability vector includes: Multiple Gaussian distributions are represented by the predicted mean vector and the predicted standard deviation vector to obtain multiple candidate Gaussian distributions. Based on the predicted probability vector, multiple candidate Gaussian distributions are sampled and predicted to obtain the predicted movement position of the target object within the second target time period in the target game. The predicted movement trajectory is determined based on the predicted movement location.

3. The method according to claim 2, characterized in that, The step of performing multiple Gaussian distribution representations based on the predicted mean vector and the predicted standard deviation vector to obtain multiple candidate Gaussian distributions includes: Multiple predicted means are determined based on the predicted mean vector; Multiple prediction covariance matrices are determined based on the prediction standard deviation vector; Multiple Gaussian distributions are applied to the multiple predicted means and multiple predicted covariance matrices to obtain multiple candidate Gaussian distributions.

4. The method according to claim 2, characterized in that, The step of sampling and predicting multiple candidate Gaussian distributions based on the predicted probability vector to obtain the predicted movement position of the target object within the second target time period in the target game includes: The predicted probability vector is weighted to obtain the predicted weights for each candidate Gaussian distribution; The target Gaussian distribution is determined from the multiple candidate Gaussian distributions based on the multiple prediction weights corresponding to the multiple candidate Gaussian distributions; The predicted movement position is obtained by sampling and predicting the target Gaussian distribution.

5. The method according to claim 1, characterized in that, The step of mixing and encoding the object state features, the game environment features, and the game task features to obtain the predicted output vector of the target object within the second target time period in the target game includes: The object state features, the game environment features, and the game task features are concatenated to obtain a target concatenation vector; The target concatenation vector is mixed-encoded to obtain the predicted output vector.

6. The method according to claim 5, characterized in that, The step of performing hybrid encoding on the target concatenation vector to obtain the predicted output vector includes: The target concatenation vector is subjected to self-attention-based hybrid encoding to obtain the predicted output vector.

7. The method according to claim 5, characterized in that, The step of performing hybrid encoding on the target concatenation vector to obtain the predicted output vector includes: The target concatenation vector is subjected to hybrid encoding based on multi-head attention to obtain the predicted output vector.

8. The method according to claim 5, characterized in that, The step of concatenating the object state features, the game environment features, and the game task features to obtain the target concatenation vector includes: The object state features, the game environment features, and the game task features are concatenated to obtain a concatenated vector. The concatenated vector is position-encoded to obtain a position-encoded vector; The concatenated vector and the position encoding vector are concatenated to obtain the target concatenated vector.

9. The method according to any one of claims 1-8, characterized in that, The target game is a game in progress, and the object state information includes the first object state information of the target object and the second object state information of the game object.

10. The method according to any one of claims 1-8, characterized in that, The method further includes: The sample object state information, sample game environment information, and sample game task information associated with the sample object within the first historical time period in the sample game are represented by features to obtain sample object state features, sample game environment features, and sample game task features; the sample object state information includes sample movement state information; Using an initial encoding model, the sample object state features, the sample game environment features, and the sample game task features are mixed-encoded to obtain the sample prediction output vector of the sample object in the second historical time period of the sample game; the second historical time period is after the first historical time period. For each of the sample prediction output vectors, the sample prediction output vectors are decomposed to obtain the sample prediction mean vector, sample prediction standard deviation vector and sample prediction probability vector of the initial Gaussian model. Based on the sample prediction mean vector, the sample prediction standard deviation vector, and the sample prediction probability vector, the predicted probability density of the movement position of the sample object within the second historical time period in the sample game is obtained; Based on the predicted probability density and the actual movement trajectory of the sample object within the second historical time period in the sample game, the initial encoding model and the initial Gaussian model are trained to obtain a hybrid encoding model and a Gaussian mixture model for movement trajectory prediction.

11. The method according to claim 10, characterized in that, The step of obtaining the predicted probability density of the movement position of the sample object within the second historical time period in the sample game based on the sample predicted mean vector, the sample predicted standard deviation vector, and the sample predicted probability vector includes: Using the initial Gaussian model, multiple Gaussian distributions are obtained by representing the sample prediction mean vector and the sample prediction standard deviation vector with multiple Gaussian distributions. The sample prediction probability vector is weighted and distributed to obtain the Gaussian distribution sample weights for each sample; The predicted probability density is obtained by weighting multiple Gaussian distributions of the samples and the corresponding sample weights.

12. The method according to claim 10, characterized in that, The step of training the initial encoding model and the initial Gaussian model based on the predicted probability density and the actual movement trajectory of the sample object within the second historical time period in the sample game, to obtain a hybrid encoding model and a Gaussian mixture model for movement trajectory prediction, includes: Determine the predicted probability density on the actual movement trajectory; By maximizing each of the predicted probabilities, the initial encoding model and the initial Gaussian model are trained to obtain the mixed encoding model and the Gaussian mixture model.

13. The method according to claim 1, characterized in that, The method further includes: Based on the predicted movement trajectory, the target object is subjected to downstream applications based on the movement trajectory to obtain downstream application results.

14. A motion trajectory prediction device, characterized in that, The device includes: a feature representation unit, a hybrid encoding unit, a vector decomposition unit, and a determination unit; The feature representation unit is used to perform feature representation on the object state information, game environment information and game task information associated with the target object within the first target time period in the target game, and obtain object state features, game environment features and game task features; the object state information includes movement trajectory information. The hybrid encoding unit is used to perform hybrid encoding on the object state features, the game environment features, and the game task features to obtain the predicted output vector of the target object in the second target time period in the target game; the second target time period is after the first target time period. The vector decomposition unit is used to perform vector decomposition on each predicted output vector to obtain a predicted mean vector, a predicted standard deviation vector, and a predicted probability vector. The determining unit is used to determine the predicted movement trajectory of the target object within the second target time period in the target game based on the predicted mean vector, the predicted standard deviation vector, and the predicted probability vector.

15. A computer device, characterized in that, The computer device includes a processor and memory: The memory is used to store computer programs and to transfer the computer programs to the processor; The processor is configured to execute the method according to any one of claims 1-13 according to instructions in the computer program.

16. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store a computer program that, when run on a computer device, causes the computer device to perform the method according to any one of claims 1-13.

17. A computer program product, comprising a computer program, characterized in that, When the computer program is run on a computer device, it causes the computer device to perform the method according to any one of claims 1-13.

Images