Simulated and situated models of chemical trail following in ants

A simple neural network model for ant chemical trail following is proposed. It performs temporal differencing and comparison to control an osmotropotaxis response. The mechanism was first tested by implementing it in a virtual simulation, and using a genetic algorithm to find appropriate connection strengths. Resulting behavioural measures show strong similarity to data from ants and previous algorithmic models, including: non-linear effects of varying chemical strength; failure to follow trails crossed at large angles; worse trail following at faster speeds and better trail following with longer antennae. In a realworld implementation using chemical sensors on a robot following an alcohol-based trail, it was found necessary to use a somewhat different set of weightings to cope with the inherent unreliability of detecting chemical concentrations. It was still possible to show qualitatively similar behaviour under the same experimental conditions as the simulation model. We argue that these results may illustrate the nature of the agent-environment task space rather than proving the model ‘correct’; but that the model nevertheless provides useful pointers to further investigation of this biological system.