Transformer Model for Game Action Prediction via Text Conversion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing technologies face challenges in utilizing neural network technology for learning game history data to predict human-like actions in online games, particularly in turn-based battle games, due to difficulties in processing sequential data and contextual relationships.
Innovation Solution
A method is developed to generate a trained model for predicting user actions by converting game state and action data into text data, using a deep learning model that determines weights based on user information and generates training data with varied sentence orders, allowing for natural language processing and sequential data learning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If transformer neural network technology is used to learn causal relationships or order relationships in turn-based battle games, then the ability to recognize context is improved, but it has been difficult to use this technology for learning game history data
Solution Approach 1:
The patent introduces text data as an intermediary representation layer between game history data and the transformer neural network. Game states and actions are converted into text formats (game state text and action text), which can then be processed by the pre-trained natural language model. This intermediary text representation enables the transformer network to effectively learn from game history data while maintaining its context recognition capabilities.
Solution Approach 2:
The patent changes the parameter representation of game data by converting structured game state data and action data into text-based parameters. This transformation allows the transformer neural network, originally designed for natural language processing, to process game history data effectively. The text representation includes varying sentence structures and orders that capture the essence of game states and actions.
2Adaptability or versatility
If game state data and action data are converted into text data for natural language processing, then the ability to use transformer neural networks is improved, but the complexity of data processing increases
Solution Approach 1:
The patent segments the data processing into distinct modules: a game state text generation unit that converts game state data into text, an action text generation unit that converts action data into text, and a training data generation unit that combines these texts. This segmentation manages complexity by dividing the conversion process into manageable, specialized components rather than attempting a monolithic transformation.
Solution Approach 2:
The patent employs a pre-trained natural language model (such as BERT or GPT) that has already learned linguistic patterns and contextual relationships before being applied to game data. This preliminary training of the model on general language data reduces the complexity of training from scratch on game data, as the model already possesses foundational language understanding capabilities that can be fine-tuned for game-specific tasks.
3Productivity
If multiple items of game state text with different orders are generated from one game state, then the training data diversity is improved, but the data generation time increases
Solution Approach 1:
The patent generates multiple variations of game state text by periodically applying different sentence ordering patterns or transformations to the same underlying game state data. Instead of generating all possible permutations, the system applies a limited set of periodic transformations (such as changing subject-verb-object order or rephrasing certain elements) to create diverse training examples efficiently, balancing diversity with generation time.
Data Source
AI summary
One or more embodiments of the invention is a method for generating a trained model for predicting an action to be selected by a user in a game that proceeds in accordance with actions selected by the user, while updating game states, the method including: determining weights for individual history-data element groups; generating training data from data of game states and actions included in the history-data element groups; and generating a trained model on the basis of the generated training data, wherein the generation of training data includes generating a number of items of game state text as game state text corresponding to one game state, having different orders of a plurality of text elements, the number being based on the determined weight, and generating training data including pairs of the individual generated items of game state text and corresponding action text.