Model-free guiding of Boolean control networks: Reinforcement learning and adversarial optimization

Abstract

Guiding Boolean control networks (BCNs) toward desired states via control strategies is essential in practical applications. However, the model-driven paradigm faces limitations in adaptability and flexibility due to system complexity and uncertainty. This paper introduces a novel framework based on generative adversarial networks (GANs) that combines reinforcement learning (RL) and Markov decision process (FBCN_RM) to formulate control strategies over time series. Introducing maximum likelihood estimation (MLE) to handle incomplete state sequences and employing Policy Gradient (PG) for reward assessment to estimate the potential maximum reward between complex conditions and agents. During the adversarial process, control strategy is generated by GANs with state nodes inferred from model-based approaches. Furthermore, a novel interpretable prior knowledge is introduced to achieve higher accuracy and generalization in building near-truest strategy. Finally, the effectiveness of our approach is validated through two examples.