Learning the rules of a game: neural conditioning in human-robot interaction with delayed rewards

Learning in human-robot interaction, as well as in human-to-human situations, is characterised by noisy stimuli, variable timing of stimuli and actions, and delayed rewards. A recent model of neural learning, based on modulated plasticity, suggested the use of rare correlations and eligibility traces to model conditioning in real-world situations with uncertain timing. The current study tests neural learning with rare correlations in a human-robot realistic teaching scenario. The humanoid robot iCub learns the rules of the game rock-paper-scissors while playing with a human tutor. The feedback of the tutor is often delayed, missing, or at times even incorrect. Nevertheless, the neural system learns with great robustness and similar performance both in simulation and in robotic experiments. The results demonstrate the efficacy of the plasticity rule based on rare correlations in implementing robotic neural conditioning.