Project Overview
Boolean network (BN) and its extension probabilistic Boolean network (PBN) are popular mathematical models for studying genetic regulatory networks. Apart from applications in genetic networks, BNs and PBNs also find many other applications in modeling financial risk, manufacturing systems and service systems.
In this project, we propose a novel algorithm based on a reinforcement learning model for constructing sparse PBNs (probabilistic Boolean networks) from practical data to an algorithm to investigate the relationship between different states. The main idea is to decompose the PBN into several BN matrices which contain information about the attractor cycles. We applied the proposed methods to modeling correlated defaults and nursing home management.
I am responsible for designing the algorithm to decompose the PBN to BNs while the entropy should be maximized. The solution space is exponentially large. So we decided to adopt the Reinforcement Learning approach to this problem. However, the action space is still exponentially large. Hence, I modified the RL model to reduce the action space such that the model can converge. Then, I applied Deep Deterministic Policy Gradient (DDPG) to this problem.