Episodic Exploration for Deep Deterministic Policies: An Application to StarCraft Micromanagement Tasks

TL;DR

This paper introduces a reinforcement learning approach with deep neural networks and a novel exploration algorithm to master micromanagement tasks in StarCraft, demonstrating success in complex, large-scale scenarios.

Contribution

It presents a new heuristic reinforcement learning algorithm combining policy exploration and backpropagation, tailored for large state-action spaces in real-time strategy games.

Findings

Successfully learned strategies for up to 15 agents.

Outperformed Q-learning and REINFORCE in these tasks.

Demonstrated the effectiveness of direct policy exploration.

Abstract

We consider scenarios from the real-time strategy game StarCraft as new benchmarks for reinforcement learning algorithms. We propose micromanagement tasks, which present the problem of the short-term, low-level control of army members during a battle. From a reinforcement learning point of view, these scenarios are challenging because the state-action space is very large, and because there is no obvious feature representation for the state-action evaluation function. We describe our approach to tackle the micromanagement scenarios with deep neural network controllers from raw state features given by the game engine. In addition, we present a heuristic reinforcement learning algorithm which combines direct exploration in the policy space and backpropagation. This algorithm allows for the collection of traces for learning using deterministic policies, which appears much more efficient than, for example, {\epsilon}-greedy exploration. Experiments show that with this algorithm, we successfully learn non-trivial strategies for scenarios with armies of up to 15 agents, where both Q-learning and REINFORCE struggle.