Evolving Memory in Gene Regulatory Networks for Artificial Growth

Using simple artificial evolution to assist in the design of complex engineering systems has a number of issues regarding diversity, scalability etc. A possible improvement is to look again towards biological systems and consider both evolution and development processes. Artificial evolutionary development (Evo-Devo) systems often lack intrinsic mechanisms to regulate the termination of artificial growth. Biological systems, from which the analogy is drawn, make use of internal and external states—the organism’s makeup and environmental forces. This perpetual state can act as a form of memory within the system, allowing for the build up of chemical concentrations and gradients, for example. This work explores how the introduction of memory, through the support of an LSTM-Gene Regulatory Network (LSTM-GRN), allows for the emergence of stopping criteria in an Evo-Devo design system. The results show that memory evolves such that it is able to control the development process and its termination in simpler settings, although this behaviour is less consistent in more complex environments. The LSTM-GRN is capable of performing well in simple tasks such as linear space exploration but is constrained by its fixed topology when applied to more complex branching exploration tasks. This work sets the foundation for further exploration into how memory can impact the diversity of solutions found in Evo-Devo systems, thus providing more varied approaches to tackling more complex design challenges. Where more complex design requirements are necessary, designers can rely on more automated processes to search a design space. With more nuanced stopping criteria, this space becomes much richer, allowing for better design identification.