Estimation of Human Meta-Strategies through Comparative Analysis of Behavioral Transition Patterns in Collision Avoidance with Agents Employing Active and Passive Strategies

Abstract

In cooperative collision avoidance tasks, individuals may adopt not only passive strategies― such as avoiding the other based on their trajectory―but also active strategies that involve proactively indicating their own path to influence the other’s behavior. For robots to switch between such strategies, they must possess meta-strategies and estimate the counterpart’s meta-strategy based on behavioral transition patterns. This study investigates human-agent interaction in a passing scenario, where agents employ either active or passive strategies. By comparing the behavioral transition patterns of participants with those of strategic agents, we classify human behavioral tendencies. In our previous research, we analyzed participant behavior by comparing the frequency and appropriateness of specific actions in response to given situations and strategies. In this work, we extend that approach by extracting each participant ’ s trajectory between outer and inner paths as sequential data, representing them as tree structures. We then introduce a similarity score based on the length and frequency of shared prefixes within these trees. This score enables the visualization and classification of similarities in behavioral tendencies across participants using heatmaps. The resulting participant clusters were found to be consistent with strategy-based groupings identified in our earlier research.