Discovering a set of policies for the worst case reward
This work is significant for researchers and practitioners in reinforcement learning who need to develop robust policy sets capable of handling a variety of task variations, especially when worst-case performance is a critical concern.
This paper addresses the challenge of creating a set of policies that can be combined to solve multiple reinforcement learning tasks, each defined by a linear combination of features. The authors propose a policy iteration algorithm that builds a set of policies to maximize the worst-case performance of a 'set-max policy' (SMP) across these tasks, demonstrating strict performance improvement at each iteration.
We study the problem of how to construct a set of policies that can be composed together to solve a collection of reinforcement learning tasks. Each task is a different reward function defined as a linear combination of known features. We consider a specific class of policy compositions which we call set improving policies (SIPs): given a set of policies and a set of tasks, a SIP is any composition of the former whose performance is at least as good as that of its constituents across all the tasks. We focus on the most conservative instantiation of SIPs, set-max policies (SMPs), so our analysis extends to any SIP. This includes known policy-composition operators like generalized policy improvement. Our main contribution is a policy iteration algorithm that builds a set of policies in order to maximize the worst-case performance of the resulting SMP on the set of tasks. The algorithm works by successively adding new policies to the set. We show that the worst-case performance of the resulting SMP strictly improves at each iteration, and the algorithm only stops when there does not exist a policy that leads to improved performance. We empirically evaluate our algorithm on a grid world and also on a set of domains from the DeepMind control suite. We confirm our theoretical results regarding the monotonically improving performance of our algorithm. Interestingly, we also show empirically that the sets of policies computed by the algorithm are diverse, leading to different trajectories in the grid world and very distinct locomotion skills in the control suite.