Work

Tool Performance Balancing and Behavior Modeling in a Large-Scale Competitive Virtual Environment : An Experimental Platform for AI-Assisted Intervention Design

Year: 2026
Type: article
Abstract: This paper investigates how adjustments to tool parameters affect agent behavior and outcome distributions in a large-scale competitive virtual environment inspired by digital last-survivor competition formats.We define "tools" as classes of inenvironment items with distinct performance characteristics and formulate a balancing problem in which the win rates of five tool categories are brought as close as possible to 20% each.To address this problem, we implement and compare a genetic algorithm (GA), an LLM-guided causal-reasoning approach, and hybrid optimization methods that combine the two.We further interpret the framework through Immersive Data Science (IDS), used here as a conceptual lens in which causal structures are externalized, validated, and accumulated across immersive environments rather than being compressed solely into model parameters.In the base environment, the GA achieved a win-rate variance of 0.00212, the LLM-based method achieved 0.00305, a gemma3-based hybrid achieved 0.00291, and a gpt-oss:120B-based hybrid achieved the best result, 0.00055.In the extended environments, the GA under elevation advantage achieved 0.00144, whereas the hybrid method in the scouting setting achieved 0.00085.Stored result artifacts further show that the LLM-only method briefly attained an internal best variance of 0.00044 during search but remained unstable across iterations, and that its final 100-match reevaluation yielded wide confidence intervals together with category-level differences in average successful interactions and damage.Supplementary human-play logs also exhibit markedly heterogeneous survival and interaction patterns across tool types.These results indicate that balancing is shaped not only by direct parameter adjustment, but also by the interaction among environmental structure, behavior rules, search strategy, and AI-assisted hypothesis generation.Overall, the paper presents an integrated experimental platform for intervention design in competitive multi-agent virtual environments.
Authors: Nakamura, Ryota, 696
Language: English
Open Access status: green
OpenAlex ID: W7151674357
FWCI:
Cites: 34
Cited by:
Related to: 10