Q-Prop: Sample-Efficient Policy Gradient with An Off-Policy Critic

Abstract

Model-free deep reinforcement learning (RL) methods have been successful in a wide variety of simulated domains. However, a major obstacle facing deep RL in the real world is their high sample complexity. Batch policy gradient methods offer stable learning, but at the cost of high variance, which often requires large batches. TD-style methods, such as off-policy actor-critic and Q-learning, are more sample-efficient but biased, and often require costly hyperparameter sweeps to stabilize. In this work, we aim to develop methods that combine the stability of policy gradients with the efficiency of off-policy RL. We present Q-Prop, a policy gradient method that uses a Taylor expansion of the off-policy critic as a control variate. Q-Prop is both sample efficient and stable, and effectively combines the benefits of on-policy and off-policy methods. We analyze the connection between Q-Prop and existing model-free algorithms, and use control variate theory to derive two variants of Q-Prop with conservative and aggressive adaptation. We show that conservative Q-Prop provides substantial gains in sample efficiency over trust region policy optimization (TRPO) with generalized advantage estimation (GAE), and improves stability over deep deterministic policy gradient (DDPG), the stateof-the-art on-policy and off-policy methods, on OpenAI Gym's MuJoCo continuous control environments.

Measuring Progress in Deep Reinforcement Learning Sample Efficiency

“…For the MuJoCo-gym environments, we only consider results that were reported with the v1 version of the respective environment up to 2019 as the earliest publication of the latest result we found for v1 (Abdolmaleki et al, 2018a) came out in December 2018, but include results that use v2 or an ambiguous version from 2019 and 2020. Over all, we considered TRPO (Schulman et al, 2015), DDPG (Lillicrap et al, 2015), Q-Prop (Gu et al, 2016), Soft Q-learning (Haarnoja et al, 2017), ACKTR (Wu et al, 2017), PPO (Schulman et al, 2017), Clipped Action Policy Gradients (Fujita & Maeda, 2018), TD3 (Fujimoto et al, 2018), STEVE (Buckman et al, 2018), SAC (Haarnoja et al, 2018) and Relative Entropy Regularized Policy Iteration (Abdolmaleki et al, 2018a) for gym-v1.…”

Deep Reinforcement Learning for Conservation Decisions

“…The amount of time needed for an agent to learn high reward yielding behavior cannot be predetermined and depends on a host of factors including the complexity of the environment, the complexity of the agent, and more. Yet, overall, it has been well established that deep RL agents tend to be very sample inefficient (Gu et al 2017), so it is recommended to provide a generous training budget for these agents.…”

On Proximal Policy Optimization's Heavy-tailed Gradients

“…Medina & Yang (2016) extended this approach to the problem of linear bandits under heavy-tailed noise. There is also a long line of work in deep RL which focuses on reducing the variance of stochastic policy gradients (Gu et al, 2016;Wu et al, 2018;Cheng et al, 2020). On the ip side, Chung et al (2020) highlighted the bene cial impacts of stochasticity of policy gradients on the optimization process.…”