Multi-dimensional feature fusion labeling method based on chess playing stage

Abstract

The invention relates to the technical field of artificial intelligence and Go data processing, in particular to a multi-dimensional feature fusion labeling method based on a chess playing stage, and discloses an intelligent labeling method aiming at Go data set construction. Chess manual data is dynamically divided into three stages through chess stage recognition; differentiated feature labeling strategies are adopted for each stage, meanwhile, chess player style features are creatively fused, and a multi-dimensional labeling system with time-space correlation characteristics is formed. According to the marking method, the training efficiency and the decision-making quality of the Go AI during strategy optimization are remarkably improved by constructing the three-dimensional feature marking matrix with the time sequence and the marking information.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and Go data processing technology, and more specifically, to a multi-dimensional feature fusion labeling method based on the game stage of Go. Background Art

[0002] Data labeling is a crucial task in Go AI training. Both the AI's training speed and game performance rely on high-quality labeled data. The quality of labeled data directly impacts the model's performance and accuracy. Therefore, accurate labeling of Go data is considered fundamental to building high-performance models.

[0003] In Go AI training, existing data annotation methods typically use a unified strategy, failing to distinguish between the characteristics of the opening, middle game, and endgame stages, and lacking modeling of player style. For example, traditional methods only annotate the names of fixed patterns or the results of wins and losses, failing to capture key decision-making patterns at different stages (such as the transition between offense and defense in the middle game and the calculation of the number of points in the endgame), making it difficult for AI to learn strategic logic that is spatially and temporally correlated.

[0004] Because Go is a complex intellectual game, different stages require different strategies and decision-making emphases. Therefore, the annotation of Go datasets should be interpretable to improve the consistency and accuracy of annotation results, thereby optimizing the effectiveness of Go AI training. Therefore, to address the strategic differences between different Go stages, it is urgent to develop a annotation method that integrates multi-dimensional features to improve the training efficiency and decision-making quality of Go AI during strategy optimization. Summary of the Invention

[0005] In response to the shortcomings of the existing technology, the purpose of the present invention is to provide a multi-dimensional feature fusion annotation method based on the chess game stage. By fusing and annotating the multi-dimensional features of different game stages, the accuracy and comprehensiveness of chess game annotation are improved, providing better data support for chess AI training, chess game analysis, and chess teaching.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] A multi-dimensional feature fusion annotation method based on the chess game stage mainly includes the following processes:

[0008] Step 1: Identify the chess game stages.

[0009] Training a Go game stage recognition model: Go game records are collected from professional Go player games, AI game data, and Go game platform data. The collected records are preprocessed and converted into two-dimensional grayscale images, where each pixel represents a board coordinate, with black, white, and empty spaces corresponding to different grayscale values ​​(1, -1, 0). These serve as the model's input tensor, removing noise and incomplete data. The preprocessed record data is divided into training and test sets based on a specific ratio. The game stage recognition model is trained using the training set. An appropriate deep learning model, such as a convolutional neural network (CNN), is selected. Parameters are optimized to enable the model to extract spatial features on the board, identify patterns at different stages, and accurately distinguish between the opening, middlegame, and endgame phases. The trained model is evaluated on the test set to ensure that its accuracy meets the set threshold.

[0010] Game phase recognition: Use the game phase recognition model to identify the phases of unlabeled chess records, obtain the timestamps of the opening phase, middle game phase, and endgame phase, and convert them into the number of moves in the corresponding phase.

[0011] Step 2: Differentiation feature annotation.

[0012] (a) Annotation strategy in the layout phase.

[0013] Construct a fixed pattern library (fixed pattern: a classic local move sequence verified by long-term practice in Go, with a fixed order and coordinate pattern): For an unlabeled game record data, convert the SGF (Smart Game Format, a standard file format for recording Go games) game record into a time-position coordinate sequence S = {(t n ,x n ,y n )}, where t n Indicates the time of the move (numbering starts from 1), (x n ,y n ) represents the coordinates of the nth step on the board. Based on known experience and knowledge, common pattern templates are summarized and sorted out (the existing pattern library on the Go game platform can be used as a reference) j ={(t 1j ,x 1j ,y 1j ),(t 2j ,x 2j ,y 2j ),...(t kj ,x kj ,y kj )}, where k is the number of steps of the pattern, and the pattern library D = {P1, P2, ..., P m}, m is the total number of patterns included in the pattern library;

[0014] Set matching rules: Set identical matching rules and similar matching rules to accurately identify the use of fixed patterns in unlabeled chess records;

[0015] Feature annotation of the pattern: traverse each pattern P in the pattern library D j ; For each formula P j , starting from the starting position 1 in the chess record sequence S, try to match in sequence until all the moves in the layout phase (obtained in step 1) are traversed; first perform an identical match, then perform a similar match judgment; after traversing all the patterns in the pattern library, output all the matched patterns and mark the corresponding pattern in the corresponding sequence of the chess record;

[0016] (b) Marking strategy in the middle game.

[0017] Marking key decision points: The middle game is a complex and dynamic stage that determines the outcome of a game. Marking key decision points is crucial for AI to predict player behavior and win rates, as well as to make high-quality decisions during the game. A policy network is used to mark win rate predictions. By setting a win rate change threshold, key decision points are marked based on the change in win rate (if it exceeds the threshold).

[0018] Decision content extraction and annotation: Use keywords from the video or text commentary that matches the chess game to annotate key decision points.

[0019] (c) Endgame marking strategy.

[0020] Marking the value of point difference (point difference: the difference in point count between the two sides in a local area during the endgame phase, reflecting the impact of the current move on the final outcome): During the endgame phase of the game record obtained in step 1, the board is divided into n independent areas, which meet the following conditions: a local area is surrounded by live pieces that both sides have determined cannot be changed (for example, the open space is surrounded by their own live pieces, which the opponent cannot break), and the change in the point count (number of empty spaces) in this area is only related to the moves made by the two sides within this area, without considering the moves outside. For each area, calculate the point difference of the first and second moves of the black and white chess players, and mark the point difference value of each area;

[0021] Optimal Closing Strategy Annotation: This system builds a state transition diagram model and uses mathematical programming to quantitatively analyze the gains and losses of different moves. This allows for a precise depiction of the dynamic evolution of the game during the closing phase, thereby determining the optimal closing strategy. This comprehensive annotation of the endgame helps Go AI master the detailed calculations and strategy optimization required for the closing phase, resulting in higher-quality closing performance in actual games.

[0022] Step 3: Mark the chess player’s style.

[0023] Accurately classifying chess player styles can significantly improve AI's ability to predict moves against different opponents and increase its adaptability to opponents with different styles.

[0024] Step 4: Construct a three-dimensional feature annotation matrix.

[0025] Game stage dimension: Arrange the opening, middle game, and endgame stages in sequence to build a temporal context for the game record, making it easier for the model to capture the changing patterns of strategies at different stages;

[0026] Refine the dimension of annotation information: in the layout phase, mark the usage of the pattern and define the pattern encoding set D = {d1, d2, ..., d n}, using one-hot encoding vector e d ∈R n Indicates that if the formula d is used j , then e d (j) = 1, the rest are 0; in the middle stage, key decision points and decision contents are marked, and the key point type set K = {k1, k2, ..., k m}, also using one-hot encoding e k ∈R m , the decision content is encoded into c∈R through text-to-vector technology (such as Word2Vec) p ; In the endgame, the local difference is marked and normalized, and the optimal closing strategy M in step 2 is defined as M = {m1, m2, ..., m q}, where m t =(x t ,y t ) is the normalized coordinate of the t-th move relative to the center of the chessboard (10,10).

[0027] Player style dimension: define the style vector s∈R M , one dimension represents the radical-conservative tendency, and the other dimension represents the field-external preference (field: the actual occupied territory; external: the development potential and control power of the peripheral territory), and the value range is [-1, 1]. Flatten s into a one-dimensional vector, and the final depth dimension feature vector is: f = vec(s)∈R M .

[0028] The resulting three-dimensional feature annotation matrix uses game stages as the vertical dimension, arranging the opening, middlegame, and endgame stages in order. The horizontal dimension uses annotation information to cover the use of fixed opening patterns, key middlegame decision points and content, the value of endgame point differences, and optimal closing strategies. The depth dimension uses the feature vectors of player styles to enhance the Go AI's predictive capabilities for players of different styles. This multi-dimensional fusion provides richer and more accurate feature data for Go AI training, enabling the AI ​​to more comprehensively understand the logic of game development and improve its decision-making and game performance.

[0029] Furthermore, in step 2, the matching rule setting for marking the usage of the pattern in the layout phase includes the following steps:

[0030] First, for the chess sequence S and the pattern P j , if there is a starting position s(1≤s≤n-k+1) such that the k elements starting from s are consistent with the formula P j All elements of the game have the same order of placement time and coordinates, that is: (t s+i-1, x s+i-1 ,y s+i-1 )=(t ij ,x ij ,y ij ), where i = 1, 2, ... k, then it is determined that the chess record uses the fixed pattern P at this position. j Secondly, in order to handle some subtle differences, a similar matching rule is defined. If there is a starting position s (1≤s≤n-k+1), the following conditions are met:

[0031] 1. The order of placing the pieces is the same, i.e. t s+i-1 =t ij (i=1,2,...,k);

[0032] 2. For coordinates (x s+i-1 ,y s+i-1 ), calculate to satisfy The number of elements of j If the ratio of the total number of moves exceeds the similarity ratio threshold, it is determined that the chess record uses the fixed pattern P at this position. j , where d is the set distance threshold, indicating that the coordinates are allowed to differ by d units.

[0033] Furthermore, in step 2, we mark the key decision points in the middle game. The specific method is as follows:

[0034] A deep convolutional neural network is used as the basic architecture of the policy network, combined with a residual network structure to improve network training efficiency and performance. The network input is the chessboard state tensor B (including the sequence of moves, the state of the Qi around the pieces, the player's turn rights, forbidden points, etc.). It is processed through multiple layers of convolutional layers, pooling layers, and residual blocks. Finally, a fully connected layer outputs the predicted win rate value p∈[0,1], that is, p=Softmax(f(B)), where f represents a network function with parameters. The Softmax function is used to convert the output into a probability distribution. The cross-entropy loss function is used to measure the difference between the predicted win rate p and the true label y. The network parameters are updated through stochastic gradient descent to minimize the loss function and complete the training of the policy network.

[0035] Analyze the game process of a chess game. After each move in the middle game (obtained from step 1), the current chessboard state B is t (t represents the current number of hands) is input into the trained strategy network to obtain the predicted winning rate value p under this situation. t , that is, p t =Softmax(f(B t ), the predicted value sequence of the winning rate of the middle game process is recorded as {p1,p2,...,p T}, where T is the total number of moves played in the middle game;

[0036] In order to identify key decision points, calculate the win rate fluctuation value Δp after each move t , the calculation formula is as follows:

[0037] Δp t =|p t -p t-1 |;

[0038] Among them, p0 is the initial winning rate, and the default value is 0.5. For the winning rate fluctuation threshold ε, when the winning rate fluctuation value of a certain move is Δp t >ε, the move is considered a key decision point, where ε is twice the standard deviation of the fluctuation value of the win rate of each move in the middle game. The number of moves and the board state tensor are recorded to form a key decision point set K = {B1, B2, ..., B M}, where M is the number of key decision points.

[0039] Furthermore, the decision content extraction and annotation in step 2 is as follows:

[0040] Using video parsing tools (such as FFmpeg), unlabeled commentary videos of chess records are broken down into image sequences at a fixed frame rate. Optical Character Recognition (OCR) technology is then used to identify the placement and time of chess pieces on the board. The audio commentary in the video is then extracted and converted into text using speech-to-text technology. For text commentary on chess record websites and professional tournament reports, move number tags are matched to parse out key terms, tactical analysis, and other content. A correspondence between commentary content and move numbers is established, and the video and text commentary data are linked using timestamps or move number numbers. Natural language processing technology is used to extract decision-related keywords from the commentary content, such as corner hanging, flanking, and side breaking. The extracted decision content is then matched with the move numbers in the set K of key decision points, and each key decision point is labeled with its corresponding decision content.

[0041] Furthermore, the specific method for marking the optimal closing strategy in step 2 is as follows:

[0042] By constructing a state transition graph model, with the number of points on the chessboard as the graph node, a quantitative model is constructed to numerically represent the actual points (the number of intersections that one side's pieces have actually occupied), the potential points (based on the current situation and subsequent development possibilities, the points that one side has a high chance of converting into its own territory in the future), and the value of the chess pattern. Directed edges are used to represent the changes in points caused by moves. Mathematical programming is used to quantitatively analyze the gains and losses in points of different moves, achieving a precise description of the dynamic evolution of the chess game in the final stage.

[0043] Select appropriate algorithms, such as backtracking, dynamic programming, and depth-first search, to systematically traverse the state transition graph. As the traversal progresses, the algorithm constructs a sequence of optimal closing strategies from the current game state to the final game, which it uses as annotations for the endgame phase of the game. This annotation not only meets the requirements for optimality of local decisions but also ensures the optimal solution for the closing strategy at the global level, ultimately maximizing the game's point gain.

[0044] Furthermore, the specific process of player style annotation in step 3 is as follows:

[0045] To categorize chess styles, we selected several offensive, defensive, and balanced players, collected 500 games from each player, and labeled the game records according to their style. Experiments have shown that certain statistical characteristics of chess playing by players with different styles differ significantly. We then analyzed these game records for features, including global statistical features, local chess pattern features, and temporal dynamic features. Global statistical features include attack frequency (the frequency of offensive moves), ko fights (a seesaw battle between two sides for territory), and win rate fluctuations (offensive players have greater fluctuations). Local chess style features include the number of pieces with higher energy (defensive players have higher average energy than offensive players), pattern selection tendencies (different patterns correspond to different chess style preferences), and temporal dynamic features (the percentage of moves selecting popular positions in the layout phase, the number of offensive and defensive transitions in the middle game phase, and the degree of fluctuation in the number of points in the endgame phase). Continuous feature data was normalized. When constructing the SVM model, the RBF kernel was selected to handle nonlinear relationships, and the one-to-many classification strategy was selected. During model training, grid search was used, and the optimal parameters were selected through cross-validation. The trained SVM model was used to identify and annotate chess styles on unlabeled chess records.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] 1. Improved training efficiency: This invention adopts differentiated annotation strategies based on the characteristics of different stages of the Go game. The layout stage focuses on the use of fixed patterns, the middle stage focuses on the decision-making of key decision points, and the endgame stage focuses on the optimal solution of the closing strategy. This enables the Go AI to learn key information and strategies at different stages in a more targeted manner, avoids learning invalid information during the training process, improves training efficiency, and reduces the waste of training time and computing resources.

[0048] 2. Improved Decision Quality: By integrating multi-dimensional information to construct a three-dimensional feature annotation matrix, the Go AI is provided with more comprehensive and accurate feature data. By incorporating factors such as the opponent's playing style, the Go AI's adaptability to different opponents is enhanced, helping it make more reasonable and optimal decisions, thereby improving its decision quality and win rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 This is a flowchart of a multi-dimensional feature fusion annotation method based on the chess game stage. DETAILED DESCRIPTION

[0050] Example 1, refer to Figure 1 The present embodiment provides a multi-dimensional feature fusion annotation method based on the chess game stage, specifically comprising the following steps:

[0051] Step 1: Identify the chess game stages.

[0052] Training a Go game stage recognition model: Go game records are collected from professional Go player games, AI game data, and Go game platform data. The collected records are preprocessed and converted into two-dimensional grayscale images, where each pixel represents a board coordinate, with black, white, and empty spaces corresponding to different grayscale values ​​(1, -1, 0). These serve as the model's input tensor, removing noise and incomplete data. The preprocessed record data is divided into training and test sets based on a specific ratio. The game stage recognition model is trained using the training set. An appropriate deep learning model, such as a convolutional neural network (CNN), is selected. Parameters are optimized to enable the model to extract spatial features on the board, identify patterns at different stages, and accurately distinguish between the opening, middlegame, and endgame phases. The trained model is evaluated on the test set to ensure that its accuracy meets the set threshold.

[0053] Game stage identification: Use the game stage identification model to identify the stages of the chessboard, obtain the timestamps of the layout stage, middle game stage, and endgame stage, and convert them into the number of moves in the corresponding stage.

[0054] Step 2: Differentiation feature annotation.

[0055] (a) Annotation strategy in the layout phase.

[0056] Constructing a pattern library: For an unlabeled game record data, convert the SGF (Smart Game Format, a standard file format for recording Go games) game record into a time-position coordinate sequence S = {(t n ,x n ,y n )}, where t n Indicates the time of the move (numbering starts from 1), (x n ,y n ) represents the coordinates of the nth step on the board. Based on known experience and knowledge, common pattern templates are summarized and sorted out (the existing pattern library on the Go game platform can be used as a reference) j ={(t 1j ,x 1j ,y 1j ),(t 2j ,x 2j ,y 2j ),...(t kj ,x kj ,y kj )}, where k is the number of steps of the pattern, and the pattern library D = {P1, P2, ..., P m}, m is the total number of patterns included in the pattern library;

[0057] Set up matching rules: First, for the chess sequence S and the pattern P j, if there is a starting position s(1≤s≤n-k+1) such that the k elements starting from s are consistent with the formula P j All elements of the game have the same order of placement time and coordinates, that is: (t s+i-1, x s+i-1 ,y s+i-1 )=(t ij ,x ij ,y ij ), where i = 1, 2, ... k, then it is determined that the chess record uses the fixed pattern P at this position. j Secondly, in order to handle some subtle differences, a similar matching rule is defined. If there is a starting position s (1≤s≤n-k+1), the following conditions are met:

[0058] 1. The order of placing the pieces is the same, i.e. t s+i-1 =t ij (i=1,2,...,k);

[0059] 2. For coordinates (x s+i-1 ,y s+i-1 ), calculate to satisfy The number of elements of j If the ratio of the total number of moves exceeds the similarity ratio threshold, it is determined that the chess record uses the fixed pattern P at this position. j , where d is the set distance threshold, indicating that the coordinates are allowed to differ by d units. In this embodiment, for accurate recognition, the distance threshold d=1, and the similarity ratio threshold is 95%.

[0060] Feature annotation of the pattern: traverse each pattern P in the pattern library D j ; For each formula P j , starting from the starting position 1 in the chess record sequence S, try to match in sequence until all the moves in the layout phase (obtained in step 1) are traversed; first perform an identical match, then perform a similar match judgment; after traversing all the patterns in the pattern library, output all the matched patterns and mark the corresponding pattern in the corresponding sequence of the chess record;

[0061] (b) Marking strategy in the middle game.

[0062] Marking key decision points: The middle game is a complex and ever-changing stage, crucial for determining the outcome of a game. Marking key decision points is crucial for AI to predict player behavior and win rates, as well as to make high-quality decisions during the game. A strategy network is used to mark win rate predictions, and by setting a win rate change threshold, key decision points are marked based on the change in win rate (if it exceeds the threshold). The specific method is as follows:

[0063] A deep convolutional neural network is used as the basic architecture of the policy network, combined with a residual network structure to improve network training efficiency and performance. The network input is the chessboard state tensor B (including the sequence of moves, the state of the Qi around the pieces, the player's turn rights, forbidden points, etc.). It is processed through multiple layers of convolutional layers, pooling layers, and residual blocks. Finally, a fully connected layer outputs the predicted win rate value p∈[0,1], that is, p=Softmax(f(B)), where f represents a network function with parameters. The Softmax function is used to convert the output into a probability distribution. The cross-entropy loss function is used to measure the difference between the predicted win rate p and the true label y. The network parameters are updated through stochastic gradient descent to minimize the loss function and complete the training of the policy network.

[0064] Analyze the game process of a chess game. After each move in the middle game (obtained from step 1), the current chessboard state B is t (t represents the current number of hands) is input into the trained strategy network to obtain the predicted winning rate value p under this situation. t , that is, p t =Softmax(f(B t ), the predicted value sequence of the winning rate of the middle game process is recorded as {p1,p2,...,p T}, where T is the total number of moves played in the middle game;

[0065] In order to identify key decision points, calculate the win rate fluctuation value Δp after each move t , the calculation formula is as follows:

[0066] Δp t =|p t -p t-1 |;

[0067] Among them, p0 is the initial winning rate, and the default value is 0.5. For the winning rate fluctuation threshold ε, when the winning rate fluctuation value of a certain move is Δp t >ε, the move is considered a key decision point, where ε is twice the standard deviation of the win rate fluctuation value of each move in the middle game. In this embodiment, ε = 15%. The number of moves and the board state tensor are recorded to form a key decision point set K = {B1, B2, ..., B M}, where M is the number of key decision points;

[0068] Decision content extraction and annotation: Using video parsing tools (such as FFmpeg), unlabeled chess commentary videos are broken down into image sequences at a fixed frame rate. Optical Character Recognition (OCR) technology is then used to identify the placement and time of chess pieces on the board. The audio commentary in the video is then extracted and converted into text using speech-to-text technology. For text commentary on chess websites and professional tournament reports, move number tags are matched to parse out key terms, tactical analysis, and other content. A correspondence between commentary content and move numbers is established, and the video and text commentary data are linked using timestamps or move number numbers. Natural language processing technology is used to extract decision-related keywords from the commentary content, such as corner hanging, flanking, and side breaking. The extracted decision content is then matched with the move numbers in the set K of key decision points, and each key decision point is annotated with its corresponding decision content.

[0069] (c) Endgame marking strategy.

[0070] Marking the value of point difference: During the endgame phase of the game record obtained in step 1, divide the board into n independent regions. The following conditions must be met: a local region is surrounded by live stone boundaries that both players have determined cannot be changed (for example, an open space surrounded by their own live stones, which the opponent cannot break through), and the changes in the point count (number of empty spaces) in this region are only related to the moves made by both players within this region, without considering the moves outside. For each region, calculate the point difference of the first and second moves of the black and white chess players, and mark the point difference value of each region;

[0071] Optimal closing strategy annotation: To accurately solve the optimal strategy for the end of Go, a state transition graph model is constructed. With the number of points on the board as the graph node, a quantitative model is constructed to numerically represent the actual number of points (the number of intersections that one side's pieces have actually occupied), the potential number of points (based on the current situation and subsequent development possibilities, one side has a great chance of converting them into its own territory in the future) and the value of the chess pattern. Directed edges are used to represent the changes in the number of points caused by the placement of pieces. Mathematical programming is used to quantitatively analyze the gains and losses of the number of points in different ways of playing, thus achieving an accurate portrayal of the dynamic evolution of the chess game in the closing stage.

[0072] In terms of the design and implementation of the search algorithm, this embodiment uses a dynamic programming (DP) algorithm to systematically traverse the state transition diagram. The dynamic programming algorithm, with its unique optimal substructure characteristics, uses the final state of the chess game as the backtracking starting point and, through a reverse reasoning mechanism, decomposes the complex problem of solving the global optimal solution into a series of related sub-problems. As the traversal process progresses, the algorithm constructs a sequence of optimal closing strategies from the current chess game situation to the end game, and uses it as the annotation content of the chess game's endgame stage. This annotation not only meets the optimality requirements of local decision-making, but also ensures the optimal solution of the closing strategy at the global level, ultimately maximizing the number of points gained in the chess game.

[0073] Comprehensively annotate the endgame phase of the game to help Go AI master the detailed calculations and strategy optimization in the closing phase, thereby achieving higher-quality closing performance in actual games.

[0074] Step 3: Mark the chess player’s style.

[0075] Accurately classifying player styles can significantly improve AI's ability to predict moves against different opponents and increase its adaptability to opponents with different styles. The specific methods are as follows:

[0076] To categorize chess styles, we selected several offensive, defensive, and balanced players, collected 500 games from each player, and labeled the game records according to their style. Experiments have shown that certain statistical characteristics of chess playing by players with different styles differ significantly. We then analyzed these game records for features, including global statistical features, local chess pattern features, and temporal dynamic features. Among them, global statistical features include attack frequency (the frequency of using offensive moves), the frequency of ko fights (a tug-of-war between two sides for territory), and fluctuations in winning rate (offensive players have a greater fluctuation range). Local chess type features include the number of chess pieces (defensive players have a higher average number of chess pieces than offensive players), the tendency to choose set patterns (different set patterns correspond to different chess style preferences), and temporal dynamic features (the proportion of moves that choose popular positions in the layout phase, the number of offensive and defensive conversions in the middle game phase, and the degree of fluctuation in the number of points in the endgame phase). Continuous feature data are normalized. When constructing the SVM model, the RBF kernel is selected to handle nonlinear relationships, and the one-to-many classification strategy is selected. When training the model, grid search is used, and the optimal parameters are selected through cross-validation. The trained SVM model is used to identify and label the chess style of unlabeled chess records.

[0077] Step 4: Construct a three-dimensional feature annotation matrix.

[0078] Game stage dimension: Arrange the opening, middle game, and endgame stages in sequence to build a temporal context for the game record, making it easier for the model to capture the changing patterns of strategies at different stages;

[0079] Refine the dimension of annotation information: in the layout phase, mark the usage of the pattern and define the pattern encoding set D = {d1, d2, ..., d n}, using one-hot encoding vector e d ∈R n Indicates that if the formula d is used j , then e d (j) = 1, the rest are 0; in the middle stage, key decision points and decision contents are marked, and the key point type set K = {k1, k2, ..., k m}, also using one-hot encoding e k ∈R m , the decision content is encoded into c∈R through text-to-vector technology (such as Word2Vec) p ; In the endgame, the local difference is marked and normalized, and the optimal closing strategy M in step 2 is defined as M = {m1, m2, ..., m q}, where m t =(x t ,y t ) is the normalized coordinate of the t-th move relative to the center of the chessboard (10,10).

[0080] Player style dimension: define the style vector s∈R M In this embodiment, M = 2, one dimension represents the radical-conservative tendency, and the other dimension represents the field-external preference (field: the actual occupied territory; external: the development potential and control power of the peripheral territory), and the value range is [-1, 1]. After flattening s into a one-dimensional vector, the final depth dimension feature vector is: f = vec(s)∈R M .

[0081] The resulting three-dimensional feature annotation matrix uses game stages as the vertical dimension, arranging the opening, middlegame, and endgame stages in order. The horizontal dimension uses annotation information to cover the use of fixed opening patterns, key middlegame decision points and content, the value of endgame point differences, and optimal closing strategies. The depth dimension uses the feature vectors of player styles to enhance the Go AI's predictive capabilities for players of different styles. This multi-dimensional fusion provides richer and more accurate feature data for Go AI training, enabling the AI ​​to more comprehensively understand the logic of game development and improve its decision-making and game performance.

[0082] Through the detailed introduction of the above embodiments, the present invention provides a multi-dimensional feature fusion labeling method based on the game stage, which dynamically divides the game into three stages: layout, middle game and endgame, adopts differentiated feature labeling strategies for each stage, and integrates the player's style factors to form a multi-dimensional labeling system with spatiotemporal correlation, thereby improving the labeling accuracy and pertinence of the Go data set, avoiding the learning of invalid information during the AI ​​training process, and significantly improving the training efficiency and decision-making quality of the Go AI during strategy optimization.

[0083] The above formulas are all dimensionless and numerically calculated, and the preset parameters in the formulas are set by technicians in this field according to actual conditions.

[0084] The above embodiments can be implemented in whole or in part by software, hardware, firmware or any other combination. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer program are loaded or executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from one website, computer, server or data center to another website, computer, server or data center via wired or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or data center that contains one or more available media sets. The available medium can be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium. The semiconductor medium can be a solid-state drive.

[0085] It should be understood that in the various embodiments of the present application, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0086] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0087] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0088] In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0089] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0090] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

Claims

1. A multi-dimensional feature fusion annotation method based on the chess game stage, characterized by: The method flow is as follows: Step 1: Game phase identification: Use the trained game phase identification model to identify and divide the unlabeled game data into phases; Step 2: Differentiated feature annotation: Use differentiated strategies to perform feature annotation on the layout, middle game, and endgame stages of the unlabeled chess records. Step 3: Player style annotation: Identify and annotate the player's style; Step 4: Construct a three-dimensional feature annotation matrix: Construct a three-dimensional feature annotation matrix containing three dimensions: game stage, annotation information, and player style.

2. The multi-dimensional feature fusion annotation method based on the chess game stage according to claim 1 is characterized in that: The labeling strategy in the layout phase of step 2 includes the following steps: building a pattern library, setting matching rules, and labeling pattern features.

3. The multi-dimensional feature fusion annotation method based on the chess game stage according to claim 2 is characterized in that: Set matching rules as follows: First, for the chess sequence S and the pattern P j , if there is a starting position s such that the k elements starting from s are consistent with the formula P j All elements of the game have the same order of placement time and coordinates, that is: (t s+i-1, x s+i-1 ,y s+i-1 )=(t ij ,x ij ,y ij ), where i = 1, 2, ... k, then it is determined that the chess record uses the fixed pattern P at this position. j ; Secondly, define similar matching rules. If there is a starting position s, the following conditions are met: a. The order of placing the pieces is the same, i.e. t s+i-1 =t ij (i=1,2,...,k); b. For coordinates (x s+i-1 ,y s+i-1 ), calculate to satisfy The number of elements of j If the ratio of the total number of moves exceeds the similarity ratio threshold, it is determined that the chess record uses the fixed pattern P at this position. j , where d is the set distance threshold.

4. The multi-dimensional feature fusion annotation method based on the chess game stage according to claim 1 is characterized in that: In step 2, mark the key decision points in the middle game. The specific method is as follows: A deep convolutional neural network combined with a residual network structure is used as the policy network. The objective function uses the cross-entropy loss function, and the network parameters are updated through stochastic gradient descent to complete the training of the policy network. Analyze the game process of a chess game. After each move in the middle stage, the current chessboard state B t Input into the strategy network, where t represents the current number of hands, and obtain the predicted winning rate value p under this situation t , then the predicted value sequence of the winning rate of the middle game process is {p1,p2,...,p T }, where T is the total number of moves in the middle game; calculate the winning rate fluctuation value Δp after each move t , the calculation formula is as follows: Δp t =|p t -p t-1 |; Among them, p0 is the initial winning rate, the default value is 0.5; set the winning rate fluctuation threshold ε, when the winning rate fluctuation value of a certain move is Δp t >ε, the move is considered a key decision point; the move number and board state tensor are recorded to form a key decision point set K = {B1, B2, ..., B M }, where M is the number of key decision points.

5. The multi-dimensional feature fusion annotation method based on the chess game stage according to claim 1 is characterized in that: After the key decision points in the middle game phase are marked in step 2, the method for extracting and annotating the annotation content is as follows: Using video parsing tools, unlabeled commentary videos are broken down into image sequences at a fixed frame rate. Using optical character recognition (OCR) technology, the positions and times of chess pieces on the board are identified. The audio commentary is then extracted and converted into text using speech-to-text technology. For text commentary on chess websites and professional tournament reports, move number markers are matched to parse key terms, tactical analysis, and other content. A correspondence between commentary content and move numbers is established, and the video and text commentary data are linked using timestamps or move number numbers. Use natural language processing technology to extract decision-related keywords from the commentary content; correspond the extracted decision content with the number of hands in the key decision point set K, and mark the corresponding decision content for each key decision point.

6. The multi-dimensional feature fusion annotation method based on the chess game stage according to claim 1 is characterized in that: In step 2, the optimal closing strategy for the endgame is marked. The specific methods include: By constructing a state transition graph model, with the number of points on the chessboard as the graph node, a quantitative model is constructed to numerically represent the actual number of points, potential number of points, and chess pattern value. Directed edges are used to represent the changes in number of points caused by moves. Through mathematical programming, the number of points gained and lost of different moves is quantitatively analyzed, achieving a description of the dynamic evolution of the chess game in the final stage. The state transition diagram is systematically traversed. As the traversal progresses, an optimal closing strategy sequence from the current chess game situation to the end game is constructed and used as the annotation content of the endgame stage of the game.

7. The multi-dimensional feature fusion annotation method based on the chess game stage according to claim 1 is characterized in that: The specific process of player style annotation in step 3 is as follows: According to the statistical characteristics of chess records, including global statistical characteristics, local chess pattern characteristics and temporal dynamic characteristics, the chess players' styles are classified; the constructed SVM model is used to identify and annotate the chess styles of unlabeled chess record data.

8. The multi-dimensional feature fusion annotation method based on the chess game stage according to claim 1 is characterized in that: The three-dimensional feature annotation matrix in step 4 is in the following form: Game stage dimension: Arrange the opening, middle game, and endgame stages in sequence to construct the timeline of the game record; Refined annotation of information dimensions: marking the use of fixed patterns in the layout phase; marking key decision points and decision content in the middle game phase; marking the number of local point differences and the optimal closing strategy in the endgame phase; Player style dimension: define the style vector s∈R M , one dimension represents the radical-conservative tendency, and the other dimension represents the field-outside preference, both of which range from [-1, 1]. Flatten s into a one-dimensional vector, and the final feature vector of the chess player's chess style dimension is: f = vec(s)∈R M .

Images