Neural Uncertainty and Sensorimotor Robustness

Real organisms live in a world full of uncertain situations and have evolved cognitive mechanisms to cope with problems based on actions and perceptions which are not always reliable. One aspect could be related with the following questions: could neural uncertainty be beneficial from an evolutionary robotics perspective? Is uncertainty a possible mechanism for obtaining more robust artificial systems? Using the minimal cognition approach, we show that moderate levels of uncertainty in the dynamics of continuous-time recurrent networks correlates positively with behavioral robustness of the system. This correlation is possible through internal neural changes depending on the uncertainty level. We also find that controllers evolved with moderate neural uncertainty remain robust to disruptions even when uncertainty is removed during tests, suggesting that uncertainty helps evolution find regions of higher robustness in parameter space