Episodic Exploration for Deep Deterministic Policies: An Application to StarCraft Micromanagement Tasks
This addresses the problem of efficient exploration in large state-action spaces for reinforcement learning researchers, though it is incremental as it builds on existing methods for specific game scenarios.
The paper tackled the challenge of controlling multiple agents in StarCraft micromanagement tasks by proposing a heuristic reinforcement learning algorithm that combines policy space exploration and backpropagation, successfully learning non-trivial strategies for up to 15 agents where standard methods like Q-learning and REINFORCE failed.
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, ε-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.