Playing Halite IV with Deep Reinforcement Learning

Abstract: Playing games with reinforcement learning has for years been a target for research and has seen incredible breakthroughs in recent years. Reinforcement learning is a type of machine learning, which can be combined with the concept of deep learning, resulting in what is called deep reinforcement learning. The promise of deep reinforcement learning attracts businesses that aim to get an edge over traditional algorithmic methods. Our work focused on exploring the aspects of deep reinforcement learning with a DQN implementation and the game Halite IV as the environment. We created DQN agents capable of outperforming competitive solutions and tested and evaluated techniques for enhancing the DQN solution. The most insightful results include: individual decision making for a team based environment can simplify the DQN setup drastically, reward function engineering for RL is critical and a sparse reward is not practical for long time frames, self play is advantageous compared to a static opponent and low rates of exploration is beneficial in environments with built in randomness.