Q-learning with experience replay in a dynamic environment

Most research in reinforcement learning has focused on stationary environments. In this paper, we propose several adaptations of Q-learning for a dynamic environment, for both single and multiple agents. The environment consists of a grid of random rewards, where every reward is removed after a visit. We focus on experience replay, a technique that receives a lot of attention nowadays, and combine this method with Q-learning. We compare two variations of experience replay, where experiences are reused based on time or based on the obtained reward. For multi-agent reinforcement learning we compare two variations of policy representation. In the first variation the agents share a Q-function, while in the second variation both agents have a separate Q-function. Furthermore, in both variations we test the effect of reward sharing between the agents. This leads to four different multi-agent reinforcement learning algorithms, from which sharing a Q-function and sharing the rewards is the most cooperative method. The results show that in the single-agent environment both experience replay algorithms significantly outperform standard Q-learning and a greedy benchmark agent. In the multi-agent environment the highest maximum reward sum in a trial is achieved by using one Q-function and reward sharing. The highest mean reward sum is obtained with separate Q-functions and separate rewards.