Abstract
Successful movement, such as hitting a long and straight golf drive, produces a satisfying feeling. Such positive reinforcement feedback has been suggested to influence motor learning. Here, we tested the idea that the reinforcement landscape—the probability of task success given a motor action—can be manipulated to influence learning. In Experiment 1, we tested the prediction that participants experiencing a steep reinforcement landscape would learn faster than those experiencing a shallow landscape. In Experiment 2, we predicted that participants experiencing a complex landscape with multiple gradients would change where they aimed their hand such that they would ascend the steeper portion of the landscape. Participants grasped the handle of a robot arm. They reached from a home position to a displayed target. Vision of the upper limb was occluded. Critically, we shaped the reinforcement landscape by manipulating the probability of reward as a function of their angular displacement from the displayed target. Depending on the assigned landscape, participants were more likely to receive reward if they reached to the left and/or right of the displayed target. We found that participants learned at a faster rate when experiencing a steeper landscape and were more likely to ascend the steeper portion of a complex landscape. Finally, we developed a simple computational model that replicates both experiments. The model naturally reproduces several other hallmarks of human movement, such as random-walk behaviour in task-irrelevant dimensions, increased learning rates with greater movement variability and exponential learning curves.Acknowledgments: NSERC, CIHR