Robots See the Light: Directional Selection Evolves Phototactic Behavior Without Altering Brain Modularity
The capacity of a population of agents to undergo future evolutionary change is called evolvability. We use robots to test an evolvability hypothesis: selection for enhanced phototaxis will evolve modularity in the agent’s neural network. Modularity may enhance evolvability by allowing for selection to target functionally related sub-components.
We used robots called Tadros, surface-swimming autonomous agents with light sensors connected via a microcontroller running an artificial neural network (ANN) to a servo motor flapping a tail. The ANN had two inputs (intensities registered from the light sensors) and two outputs (control signals to the motor), with six hidden nodes. Each individual was tested in a light- gathering task in a circular pool with a single overhead light source. A light- sensitive photoresistor oriented vertically logged light-intensity values, which were used to calculate evolutionary fitness. We evolved ten generations of Tadros, each generation consisting of a population of ten individuals, with the first generation’s ANNs randomly generated.
The evolvable traits were the connection weights of the ANN, with possible values of +1 (excitatory), 0 (no connection), and -1 (inhibitory). The individual with the best light-gathering performance was cloned into the next generation. That individual and the other nine were then selected for reproduction with mutation, with the probability of reproduction proportional to the individual’s evolutionary fitness.
While the light-gathering ability of the population increased from generation five to six, modularity of the ANN did not, allowing us to reject the primary hypothesis. The interaction of the agent and the world is a dynamical system wherein starting conditions and seemingly minor perturbations can have large impacts on an agent’s behavior. This work was supported by the National Science Foundation.