Learning Stochastic Shortest Path with Linear Function Approximation
We study the stochastic shortest path (SSP) problem in reinforcement learning with linear function approximation, where the transition kernel is represented as a linear mixture of unknown models. We call this class of SSP problems the linear mixture SSP. We propose a novel algorithm for learning the linear mixture SSP, which can attain a Õ(d B_⋆^1.5√(K/c_min)) regret. Here K is the number of episodes, d is the dimension of the feature mapping in the mixture model, B_⋆ bounds the expected cumulative cost of the optimal policy, and c_min>0 is the lower bound of the cost function. Our algorithm also applies to the case when c_min = 0, where a Õ(K^2/3) regret is guaranteed. To the best of our knowledge, this is the first algorithm with a sublinear regret guarantee for learning linear mixture SSP. In complement to the regret upper bounds, we also prove a lower bound of Ω(d B_⋆√(K)), which nearly matches our upper bound.
READ FULL TEXT