Combining Case-Based Reasoning and Reinforcement Learning for Tactical Unit Selection in Real-Time Strategy Game AI

This paper presents a hierarchical approach to the problems inherent in parts of real-time strategy games. The overall game is decomposed into a hierarchy of sub-problems and an architecture is created that addresses a significant number of these through interconnected machine-learning (ML) techniques. Specifically, individual modules that use a combination of case-based reasoning (CBR) and reinforcement learning (RL) are organised into three distinct yet interconnected layers of reasoning. An agent is created for the RTS game StarCraft and individual modules are devised for the separate tasks that are described by the architecture. The modules are individually trained and subsequently integrated in a micromanagement agent that is evaluated in a range of test scenarios. The experimental evaluation shows that the agent is able to learn how to manage groups of units to successfully solve a number of different micromanagement scenarios.