Exploring Learning Classifier System Behaviors in Multi-Action, Turn-based Wargames
Abstract
State of the art game-playing Artificial Intelligence research focuses heavily on non-symbolic learning methods. These methods offer little explainable insight into their decision-making processes. Learning Classifier Systems (LCSs) provide an alternative. LCSs use rule-based learning, guided by a Genetic Algorithm (GA), to produce a human-readable rule-set. This thesis explores LCS usefulness in game-playing agents for multi-agent wargames. Several Multi-Agent Learning Classifier System (MALCS) variants are implemented in the wargame Stratagem MIST: a Zeroeth-Level Classifier System (ZCS), an extended Classifier System (XCS), and an Adaptive Pittsburgh Classifier System (APCS). These algorithms were tested against baseline agents as well as the Online Evolutionary Planning(OEP) algorithm. In a round-robin comparison among the agents, all LCS agents outperformed the baselines and OEP. APCS is the most effective game-playing agent while producing the most explainable output. ZCS and XCS outperformed the baselines and produced interpretable rule-sets. The results highlight the ability for symbolic methods to learn a complex wargame, outperform non-symbolic competitors, and provide replicable instructions.
Document Details
- Document Type
- Technical Report
- Publication Date
- Mar 24, 2022
- Accession Number
- AD1166929
Entities
People
- Garth J. Terlizzi
Organizations
- Air Force Institute of Technology