Deterministic Implementations for Reproducibility in Deep Reinforcement Learning
This addresses reproducibility challenges for researchers and practitioners in deep reinforcement learning, but it is incremental as it focuses on a specific algorithm and known issue.
The paper tackled the problem of nondeterminism in deep reinforcement learning training, which hinders reproducibility, by creating a deterministic implementation of deep Q-learning and measuring the impact of individual nondeterministic sources on performance variance, finding that they can substantially affect agent performance.
While deep reinforcement learning (DRL) has led to numerous successes in recent years, reproducing these successes can be extremely challenging. One reproducibility challenge particularly relevant to DRL is nondeterminism in the training process, which can substantially affect the results. Motivated by this challenge, we study the positive impacts of deterministic implementations in eliminating nondeterminism in training. To do so, we consider the particular case of the deep Q-learning algorithm, for which we produce a deterministic implementation by identifying and controlling all sources of nondeterminism in the training process. One by one, we then allow individual sources of nondeterminism to affect our otherwise deterministic implementation, and measure the impact of each source on the variance in performance. We find that individual sources of nondeterminism can substantially impact the performance of agent, illustrating the benefits of deterministic implementations. In addition, we also discuss the important role of deterministic implementations in achieving exact replicability of results.