A State-of-the-Art Review of Deep Reinforcement Learning Techniques for Real-Time Strategy Games

Abstract

Deep reinforcement learning is an active research point in the academic community now more than ever. The computational capabilities and the deep neural networks revolution can automatically find compact low-dimensional features of high-dimensional data. As the convolutional neural network gets better at dealing with the visual world, the more outstanding results show reinforcement-learning agents that can concur with various problems. Intending to develop stronger agents and go beyond the benchmark, we need environments that provide more challenge than the problems that already have solved. Real-time strategy games considered as perfect challenging environment for reinforcement learning agents because of their vast state-action spaces that mimic the state and action spaces in real-world problems. Besides, real-time strategy games are simulations with complicated interactions, so they have no cost when applying trial and error learning techniques. Deep reinforcement learning agents considered a promising step towards fully autonomous agents that learn from trial and error with little or no prior knowledge about the environment they were dealing with. The goal of this survey is to state the current state of the art of deep reinforcement learning (DRL) research in the real-time strategy game “StarCraft” introducing the creative ways in which neural networks can be used to bring us steps closer to create AI agents that can deal with the real world. Our survey will cover a theoretical background of reinforcement learning and discuss the families of reinforcement learning algorithms. We will also explain the deep reinforcement learning framework and its contributions to StarCraft. To conclude, we ask the following research question: How can deep reinforcement learning algorithms play StarCraft s full game?