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

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, 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.

18 pages, 1 figure (2 plots), 2 tables

Cited by in corpus (9)

• A Brief Survey of Deep Reinforcement Learning
• Schema Networks: Zero-shot Transfer with a Generative Causal Model of Intuitive Physics
• TorchCraft: a Library for Machine Learning Research on Real-Time Strategy Games
• ELF: An Extensive, Lightweight and Flexible Research Platform for Real-time Strategy Games
• Revisiting the Master-Slave Architecture in Multi-Agent Deep Reinforcement Learning
• STARDATA: A StarCraft AI Research Dataset
• Combining Strategic Learning and Tactical Search in Real-Time Strategy Games
• DeepConfig: Automating Data Center Network Topologies Management with Machine Learning
• Combinatorial Multi-armed Bandits for Real-Time Strategy Games