Grant Chereskin

Evolutionary Robotics uses “evolutionary algorithms” that mimic biological natural selection and survival of the fittest to optimize the wiring of robots’ artificial neural network (ANN) controllers. This is a promising, more biologically plausible way to optimize ANNs than more common methods, and research in this field can also give insight into biology. One characteristic of organisms is their impressive ability to generalize - an important and useful skill that current ANNs greatly lack. Generalization is an organism/robot’s ability to apply previous knowledge to novel situations, such as when a robot that has only ever walked on flat and sloped terrain suddenly has to walk on bumpy terrain. A robot’s failure to generalize in a world that is so dynamic and unpredictable can limit robots’ usefulness for high-stakes applications such as human rescue missions. However, there are many biological cognitive mechanisms that improve generalization and can be implemented into ANNs.

Two methods to make ANNs generalize better are (1) implementing neuroplasticity and (2) introducing variations to the conditions experienced during a robot’s evolution in a simulation. In biology, neuroplasticity is the process that allows brains’ wiring to adapt and change over time. The wiring of most ANNs used in AI can not change after initial optimization/evolution, but the wires that connect different neurons in an ANN can be “plastic” by dynamically changing their connection strengths based on nearby ‘brain’ activity in accordance to a learning rule. A learning rule is a function that describes how a connection strength changes over time, and with evolutionary robotics, learning rules’ parameters can be optimized to help robots learn to improve their performance in novel situations. On the other hand, the evolutionary algorithms that optimize robots’ ANNs can be improved to better promote generalization. One way to do this is by exposing robots’ ANNs to varying conditions during evolution, such as making ANNs control differently sized robots. However, aspects like the order in which variations are used impact this method’s effectiveness, and more broadly, the issue of a tradeoff between generalizing well and performance in known conditions is yet to be resolved.