Solving sparse reward games using deep Q-learning with demonstration and partial training
Andrew Mahisa Halim, Helena Margaretha, Kie Van Ivanky Saputra, Samuel Lukas · AIP conference proceedings · 2021
Games that has a sparse reward space are considered a challenge in the field of deep reinforcement learning. Such games are impossible to tackle with a random exploration commonly used in early stages of learning process. We present a modified Deep Q-learning architecture, as well as some small modifications to perform better in this sparse reward games. We also introduced partial training, a policy improvement technique for neural network that kickstarts an agent to get rewards faster in games with sparse rewards. We apply our methods to 2 games with varying difficulty. The results indicate that out agent learns faster and performs significantly better compared to classical Deep Q-Learning. We further shows that combined with partial training, Deep Q-learning is viable to even solve games with really sparse reward.