AI Representation Learning for Game Object Identification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In electronic games with a large number of objects, AI agents face complexity in learning and processing due to high-dimensional vectors from individually labeling each object, making it difficult to set skill effects and attack powers, and selecting appropriate characters, especially when objects have similar characteristics.
Innovation Solution
An information processing device and program that use representation learning to find characteristic vectors based on game logs, excluding information about specific objects from the game situation, allowing for efficient learning and strategy development for AI agents playing electronic games with multiple objects.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If individually labeling each object using one-hot vector, then object identification accuracy is improved, but vector dimensionality increases making processing complicated
Solution Approach 1:
The patent extracts only the necessary identifying features of objects rather than using complete one-hot vectors. By selecting and extracting key characteristics that distinguish objects, the system maintains identification accuracy while significantly reducing vector dimensionality and processing complexity.
Solution Approach 2:
Instead of treating all objects uniformly with full one-hot vectors, the patent applies local quality by using different representation strategies for different objects or contexts. Objects are represented only by the local features relevant to their identification, reducing overall system complexity while maintaining accuracy where needed.
2Reliability
If prescribing skill activation condition by game situation, then game balance is improved, but probability assessment difficulty increases
Solution Approach 1:
The patent introduces an intermediary mechanism that translates complex game situation conditions into assessable probability metrics. This intermediary layer processes the relationship between game state and skill activation, providing players with measurable probability information while maintaining balanced game design.
Solution Approach 2:
The system implements feedback mechanisms that provide players with information about skill activation probabilities based on current game situations. This feedback loop allows players to assess probabilities and make informed decisions while the system maintains game balance through controlled probability assignments.
3Measurement precision
If learning all objects with similar characteristics separately, then learning accuracy is improved, but learning efficiency decreases
Solution Approach 1:
The patent merges objects with similar characteristics into shared representation groups. By combining learning processes for objects that share common features, the system maintains accurate learning of distinguishing characteristics while significantly improving learning efficiency through shared processing pathways.
Solution Approach 2:
The patent creates universal learning mechanisms that can handle multiple objects with similar characteristics simultaneously. A single learning process serves multiple functions by processing groups of similar objects together, reducing redundant computation while maintaining the accuracy needed to distinguish between them when necessary.
Data Source
AI summary
An information processing device comprises a representation learning unit for learning characteristic vectors representing the various characteristics of objects, on the basis of a game log, which is game progress history related to an electronic game in which a plurality of objects are used and which comprises information about the game situation including information about objects that affect the game, information about objects used in said situation from among said objects, and information indicating the effect on the game arising from the use of said objects, wherein the characteristic vectors are found by performing learning using a combination of information about the effect on the game and information obtained by excluding the information about at least one of the objects from the information about the game situation.