Abstract

Although there is a long tradition of separating perception, action, and learning, these can be treated separately only under very special circumstances. There are both theoretical reasons evident from a treatment of such control tasks in the framework of Markov decision processes, as well as empirical studies having revealed this fact, especially when considering naturalistic sequential visuomotor tasks. We will present results from several studies investigating these dependencies.

First, we show that learning of representations of natural visual stimuli through generative models can explain a variety of psychophysical biases only when in addition to the statistics of the natural environment one also takes the influence of the specific imaging system such as the eye as well as the active usage of the visual system into account.

Second, we will show how human visuomotor behavior can be quantified using Bayesian inverse reinforcement learning algorithms to extract the reward functions underlying human actions. This analysis demonstrates that the guidance behavior in a navigation task does not necessarily follow the given task instructions and reveals systematic individual differences within subject’s task priorities.

Finally, we will present results from a study in which human subjects intercepted moving objects in a virtual environment. Many animals use the well studied constant bearing angle strategy, which has been characterized as a fast and frugal heuristic. The relationships governing the behavior of the environment were manipulated systematically so as to reveal that subjects can indeed quickly learn new sequential control policies. A theoretical analysis shows, that the learned behavior can only be understood by considering the observation and control uncertainties in order to successfully carry out the interception task.