Libin Liu

Libin Liu
Ph.D. Candidate (CV)
Ming Hsieh Department of Electrical and Computer Engineering
University of Southern California

Advisor: Prof. Urbashi Mitra

Sequential decision making problem can be well-modeled as a MDP or RL problem, while classical dynamic programming algorithms can be applied to obtain the optimal policy, the value function or optimal policy can be viewed as signals defined on each state with connections to other states (e.g., a Finite State Machine). On the other hand, GSP also provides an efficient representation for data in many domains. The main goal of my research is to seek reduced dimension representation of the value function/optimal policy using tools from GSP to help in efficient algorithm design.

GSP on DL (currently working on)

• Efficient DL algorithm design with GSP for structured optimal policy in RL.

GSP on RL

• Novel “image” graph representation of the optimal policy that facilitates the design policy sampling and interpolation algorithm, which achieves asymptotic zero policy error.

GSP on MDPs

• Proper subspace design using GSP for reduced dimension MDPs, perfect reconstruction of optimal policy is achieved.