DeepPursuit: Uniting Classical Wisdom and Deep RL for Sparse Recovery

In this paper, we formulate sparse signal recovery as a sequential decision making problem (modeled by Markov Decision Processes). Based on the formulation, we propose DeepPursuit, a novel sparse recovery algorithm that learns to recover sparse signals via deep reinforcement learning (RL) and Monte Carlo Tree Search (MCTS). To substantially enhance the learning speed and performance, DeepPursuit (i) employs a novel residual-type policy/value network architecture that organically incorporates the classical wisdom from the Orthogonal Matching Pursuit (OMP) algorithm, and (ii) exploits the available ground-truth knowledge to guide the MCTS during the training process. Experimental results for general random sparse signal recovery demonstrate that, with very low computational complexity, the DeepPursuit algorithm significantly outperforms the state-of-the-art algorithms. Even higher performance gains are observed with experiments on the MNIST dataset.