Lag in Situated, Embodied and Dynamical Adaptive Systems
Matthew Egbert, Andrew Keane, Claire Postlethwaite
evolutionary robotics, lag, situated embodied and dynamical robotics, dynamical systems analysis
Lag is a well recognized phenomenon among both neuroscientists and engineers, with nervous tissues, sensors, actuators and other materials often demonstrating non-negligible lagged influences. A primary focus in evolutionary robotics is the dynamical nature of cognition, but the most common approaches –e. g. evolved continuous-time recurrent neural networks (CTRNN)– employ systems of ordinary differential equations that cannot directly capture lagged influence. Engineers, control-theorists, and neuroscientists often view lag as a problem that needs to be compensated for or avoided, but in this work, we present the first stages of our investigation into how lag can be functional, i. e. can underlie ‘intelligent’ behaviours. To do so, we follow the evolutionary robotics method, using a genetic algorithm to optimise the parameters of a controller that is coupled to a simulated robot, and then performing dynamical analysis on the most successful system identified. Unlike previous ER work, the controller is modelled using delay differential equations (DDE), so as to be able to capture lagged influence. The evolved controller is highly constrained, consisting of one modeled ‘neuron’ with a single lagged recurrent connection, so as to force the system to use lag as part of its solution. The artificial evolution identifies a high-quality solution, and we present our initial dynamical systems analysis of that individual.