Decraene, James, Mahinthan Chandramohan, Fanchao Zeng, Malcolm Yoke Hean Low, and Wentong Cai. “Evolving agent-based model structures using variable-length genomes.” In Proceedings of the Fourth International workshop on Optimisation in Multi-Agent Systems , pp. 68-85. 2011.
URL1 URL2
We present a novel evolutionary computation approach to optimize agent based models using a variable-length genome representation. This evolutionary optimization technique is applied to Computational Red Teaming (CRT). CRT is a vulnerability assessment tool which was originally proposed by the military operations research community to automatically uncover critical weaknesses of operational plans. Using this agent-based simulation approach, defence analysts may subsequently examine and resolve the identied loopholes. In CRT experiments, agent-based models of simplied military scenarios are repeatedly and automatically generated, varied and executed. To date, CRT studies have used xed-length genome representation where only a xed set of agent behavioural parameters was evolved. This may prevent the generation of potentially more optimized/interesting solutions. To address this issue, we introduce the hybrid variable-length crossover to evolve the structure of agent-based models. A maritime anchorage protection scenario is examined in which the number of waypoints composing the vessel’s route is subjected to evolution. The experimental results demonstrate the eectiveness of our proposed method and suggest promising research avenues in complex agent-based model optimization.