Towards efficient exploration in Reinforcement Learning via non-reversible stochastic processes

One of the key questions in Reinforcement Learning is the compromise between exploration and exploitation. Some popular techniques to encourage exploration are the use of random noise and adding an entropy objective to the reward function. These approaches have proven useful, but suboptimal and/or costly. One key observation is that acting randomnly is not the same as having random experiences, which has lead to the Maximum Diffusion approach, with promising results. In this work, we propose using non-reversible stochastic processes, particularly PDMPs, as a framework to enforce exploration not on the reward function or through noise, but through the structure of the process.