Controlling Gene Regulatory Networks with FQI-SARSA

External control of a gene regulatory network model can help accelerate the design of treatments to make it avoid diseased states. However, inferring this model and then controlling it has a exponential complexity of time and space, making large networks inviable for model dependent approaches. This is visible in the literature as only models with at most dozens of genes could be used in control problems. We propose to apply a batch reinforcement learning method Fitted Q-Iteration Sarsa for controlling partially observable gene regulatory networks directly from data, with a new reward function and a way to create experience tuples from gene expression samples. Our framework produces approximate stochastic policies without restricting it to time series samples, allowing it to freely manage the experience tuples. Results demonstrate that our method is more effective than previous studies, with a higher shifting between undesirable to desirable states and higher expected reward.