Off-Policy Average Reward Actor-Critic with Deterministic Policy Search
This work addresses the gap in off-policy average reward algorithms for reinforcement learning, offering incremental improvements with theoretical guarantees and practical gains.
The paper tackles the understudied average reward criterion in reinforcement learning by proposing an off-policy actor-critic algorithm, achieving an ε-optimal policy with sample complexity Ω(ε^{-2.5}) and showing better empirical performance than state-of-the-art on-policy methods in MuJoCo environments.
The average reward criterion is relatively less studied as most existing works in the Reinforcement Learning literature consider the discounted reward criterion. There are few recent works that present on-policy average reward actor-critic algorithms, but average reward off-policy actor-critic is relatively less explored. In this work, we present both on-policy and off-policy deterministic policy gradient theorems for the average reward performance criterion. Using these theorems, we also present an Average Reward Off-Policy Deep Deterministic Policy Gradient (ARO-DDPG) Algorithm. We first show asymptotic convergence analysis using the ODE-based method. Subsequently, we provide a finite time analysis of the resulting stochastic approximation scheme with linear function approximator and obtain an $ε$-optimal stationary policy with a sample complexity of $Ω(ε^{-2.5})$. We compare the average reward performance of our proposed ARO-DDPG algorithm and observe better empirical performance compared to state-of-the-art on-policy average reward actor-critic algorithms over MuJoCo-based environments.