Off Environment Evaluation Using Convex Risk Minimization
This addresses the challenge of ensuring RL policy performance in real-world robotics when trained in simulation, though it is incremental as it builds on existing robust policy learning methods.
The paper tackles the problem of evaluating reinforcement learning policies in a target environment using only simulator and trajectory data, proposing a convex risk minimization algorithm to estimate model mismatch and showing it can approximate performance with a convergence rate of n^{-1/4} across simulated and real-world robotic tasks.
Applying reinforcement learning (RL) methods on robots typically involves training a policy in simulation and deploying it on a robot in the real world. Because of the model mismatch between the real world and the simulator, RL agents deployed in this manner tend to perform suboptimally. To tackle this problem, researchers have developed robust policy learning algorithms that rely on synthetic noise disturbances. However, such methods do not guarantee performance in the target environment. We propose a convex risk minimization algorithm to estimate the model mismatch between the simulator and the target domain using trajectory data from both environments. We show that this estimator can be used along with the simulator to evaluate performance of an RL agents in the target domain, effectively bridging the gap between these two environments. We also show that the convergence rate of our estimator to be of the order of ${n^{-1/4}}$, where $n$ is the number of training samples. In simulation, we demonstrate how our method effectively approximates and evaluates performance on Gridworld, Cartpole, and Reacher environments on a range of policies. We also show that the our method is able to estimate performance of a 7 DOF robotic arm using the simulator and remotely collected data from the robot in the real world.