A neurocomputational model implemented on humanoid robot for learning action selection

Computational modeling of neural circuits enhances our comprehension of brain functions. In addition to the simulation of the models which helps to anticipate cognitive processes, embodiment of these models is essential. Such embodiment would provide the setting to explain neural functioning ongoing in real environments under oncoming sensory information besides giving opportunity of implementation of intelligent systems. Even studies pursued in neuroscience seem far from achieving all these aims in intelligent systems, the pre-results using cognitive models are faster than animal experiments in leading further the understanding of cognitive processes and designing related experiments. In this study, a computational model of basal ganglia, thalamus and cortex for action selection is extended with the point neuron approach to obtain a more realistic method to investigate the model in real time task on humanoid robot platform, Darwin-Op. The spiking neural network model of cortex consists of channels for each action to be elected and plastic alI-to-alI connections from the sensory stimuli to the basal ganglia structures which are modulated with reward. In the task, the sensory inputs, namely colors, are presented to the humanoid robot and it is expected that these sensory inputs would be associated with the predefined actions by modulating the connections. Furthermore, the rearrangement of these associations with reward is performed after learning is accomplished. In this way, the embodiment of computational model provided more information on the evolution of connections through reward based learning in the action selection circuit.