Graphical Models

Biological nervous systems can deal with many real-world natural stimuli and outper-form any computer algorithm available today in most tasks involving interaction with real objects as required for example in interactive motor control problems. Moreover, they are able to learn an internal model of a controllable system solely from experience, unlike today’s most successful control algorithms. The computational architecture expressed by brains is fundamentally different from modern computer hardware, in that it is composed of billions of small computational units with information stored only in the effectiveness of the communication channels between these units. Despite this restricted form of infor-mation storage, brains are able to devise complex sequences of actions which span many time scales. We investigate neurally inspired algorithms that learn how to control systems of un-known mechanical structure solely from data collected from experience. These algorithms are designed to run on distributed computational architectures composed of many small and locally operating units. Further, we explore how motions spanning multiple tim