AdaStop: adaptive statistical testing for sound comparisons of Deep RL agents

Recently, the scientific community has questioned the statistical reproducibility of many empirical results, especially in the field of machine learning. To solve this reproducibility crisis, we propose a theoretically sound methodology to compare the overall performance of multiple algorithms with stochastic returns. We exemplify our methodology in Deep RL. Indeed, the performance of one execution of a Deep RL algorithm is random. Therefore, several independent executions are needed to accurately evaluate the overall performance. When comparing several RL algorithms, a major question is how many executions must be made and how can we ensure that the results of such a comparison are theoretically sound. When comparing several algorithms at once, the error of each comparison may accumulate and must be taken into account with a multiple tests procedure to preserve low error guarantees. We introduce AdaStop, a new statistical test based on multiple group sequential tests. When comparing algorithms, AdaStop adapts the number of executions to stop as early as possible while ensuring that we have enough information to distinguish algorithms that perform better than the others in a statistical significant way. We prove theoretically and empirically that AdaStop has a low probability of making a (family-wise) error. Finally, we illustrate the effectiveness of AdaStop in multiple Deep RL use-cases, including toy examples and challenging Mujoco environments. AdaStop is the first statistical test fitted to this sort of comparisons: AdaStop is both a significant contribution to statistics, and a major contribution to computational studies performed in reinforcement learning and in other domains. To summarize our contribution, we introduce AdaStop, a formally grounded statistical tool to let anyone answer the practical question: “Is my algorithm the new state-of-the-art?”.