Compressed kernelized least squares temporal difference learning via random projections

We propose a compressed kernelized least squares temporal difference learning (CKLSTD) algorithm for reinforcement learning in large state space by incorporate kernel trick and random projection into traditional LSTD. Our proposed method can be viewed as kernelized version of LSTD-RP and compressed version of KLSTD, wherein the proposed method preserves the merits of both of them. The experimental results suggest that our proposed method can gain better performance in high dimensional feature space.