Perceptual Metrics for Video Game Playstyle Similarity and Diversity

Chiu-Chou Lin, Wei-Chen Chiu, I-Chen Wu

Primary Area: metric learning, kernel learning, and sparse coding

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: Playstyle, Decision-Making Behavior, Metric, Similarity, Diversity, Video Game, Deep Reinforcement Learning, Human Cognition

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: New playstyle metric leverages discrete representation to measure decision-making styles in games. It utilizes human cognition concepts for similarity and quantifies AI model diversity.

Abstract: In gaming, decision-making diversity reflects the broad spectrum of styles that players can adopt. Despite the importance of this diversity, finding a universally applicable metric for it is challenging. To address this, a previous approach introduced the $\textit{Playstyle Distance}$—a method for gauging similarity between datasets using game screens and their corresponding action pairs. This method identifies comparable states in discrete representations and then computes action distribution distances. Building on it, we introduce several new techniques. These include multiscale analysis with varied state granularity, perceptual kernels rooted in psychology, and the utilization of the intersection over union method for efficient data assessment. These innovations advance playstyle measurement and offer insights into human cognition of similarity. In experiments across two racing games and seven Atari games, our metric achieves over 90\% accuracy in playstyle classification. Remarkably, this requires fewer than 512 observation-action pairs, less than half an episode in all tested games. We also develop an algorithm for assessing decision-making diversity using this metric. Our findings illuminate promising avenues for real-time game analysis and the evolution of AI with diverse playstyles.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 117