Linking total movement history to action learning
Poster presented at RLDM, Princeton, October 2013.
Abstract:
We have developed a new behavioural task which requires the agent to identify a target movement via exploratory movements. This conceptually simple, but versatile, paradigm allows the full history of movements made to be linked to learning of actions. Here we review recent results which show how the task can elucidate different aspects of the functional and anatomical basis of reinforcement learning in the human motor system. We also present a novel analysis which captures the relative influence of previous movements on learnt actions, over each point in the path of that learnt action. From a reinforcement learning perspective, this analysis can be thought of as revealing the shape of the eligibility trace: the relative strength of credit assignment to the motor efference copy across time.