Extracting Key Gene Regulatory Dynamics for the Direct Control of Mechanical Systems

Evolution produces gene regulatory networks (GRNs) able to control cells. With this inspiration we evolve artificial GRN (AGRN) genomes for the reinforcement learning control of mechanical systems with unknown dynamics, a problem domain similar in its sparse feedback to that of controlling a biological cell. From the fractal GRN (FGRN), a successful but complex GRN model, we obtain the Input-Merge-RegulateOutput (IMRO) abstraction for GRN-based controllers, in which the FGRN’s complex fractal operations are replaced by simpler ones. Computational experiments on reinforcement learning problems show significant improvements from the use of this simplified approach. We also present the first evolutionary solution to a hardened version of the acrobot problem, which previous evolutionary methods have failed on.