This work aims to model the ability of biological organisms to achieve cumulative learning, i.e. to learn increasingly more complex skills on the basis of simpler ones. In particular, we studied how a simulated kinematic robotic system composed of an arm and an eye can learn the ability to reach for an object on the basis of the ability to systematically look at the object, which, in our set-up, represented a prerequisite for the reaching task. We designed the system by following several biological constraints and investigated which kind of sub-task reinforcements might facilitate the development of the final skill. We found that the performance in the reaching task was optimized when the reinforcement signal included not only the extrinsic reinforcement provided by touching the object but also an intrinsic reinforcement given by the error in the prediction of fovea activation. We discuss how these results might explain biological data regarding the neural basis of action discovery and reinforcement learning, in particular with respect to the neuromodulator dopamine.
展开▼
