Evolving Spiking Neural Networks for Animat Foraging and Pattern Recognition

The ability to process temporal information is essential for both biological and artificial systems, especially in noisy or constantly changing environments. This work explores how evolving spiking neural network (SNN) controllers can solve nontrivial pattern recognition tasks while foraging in a 2-D world. Using a genetic algorithm and adaptive exponential integrate and fire neurones, animats evolved to respond effectively in real-time and make accurate decisions under varying levels of neural noise. The research is organised in three experimental phases: (1) Evolution of foraging animats that incorporate basic temporal signals paired with associated rewards/punishments. (2) Evolution under variable input intervals that requires some form of temporal memory. (3) Evolution with internal noise on voltage to evaluate robustness. Performance under varying conditions was evaluated by analysing the existence of neurotopological characteristics such as self-excitatory and recurrent motifs that emerged as a result of the temporal processing mechanisms of the animats. The results support the hypothesis that neuroevolution can help uncover biologically plausible topological features that increase precision in timing, decisional robustness, and generalisation. The results also highlight the extent to which internal noise in networks affects the topology that maintains optimal temporal recognition. This study helps to understand how complex temporal recognition tasks can be accomplished using SNN architectures. It also provides some methodological suggestions on robust design principles for neuromorphic and autonomous systems.