Kolloquium Dienstag, 12.07.2011, 10 Uhr, F 125

Mark Wernsdorfer, Grounding Affordances in Hierarchical Representations of Sensorimotor Interaction

Understanding human cognition implies understanding how the mind forms hierarchical representations of the world. In this project, a reinforcement learning algorithm is developed that is capable of constructing a world model, while embracing this cognitive phenomenon. Representations of concrete concepts result from basic sensorimotor optimisation and serve themselves as elements for higher order representations of abstract concepts. Sequence recognition allows for the agent to obtain a subjective notion of causality. Categorising according to causal roles eventually allows for a functional identification of representations.
Similar approaches to hierarchical reinforcement learning try to grasp functionality (encompassing e. g. context-independency or structural invariance) either by categorising patterns in the sensorimotor structure of percepts or by precategorising functional categories. In contrast, this project proposes a process of autonomous categorisation according to the functional aspects of cognitive representations, also called affordances. The intention is to ground these affordances in the interaction between agent and environment. Semantic conveyance from architect to agent will be minimised.
The agent's performance will be evaluated in a simulated environment. Motivation and realisation of this algorithm will be presented, as well as first empirical results and possible problems.