Natural intelligence in artificial creatures

What mechanisms are needed in a cognitive system, such as an animal or a robot, and how do these mechanisms interact with each other? The thesis presents a study of this problem within the field of behavior-based systems and artificial neural networks. The thesis brings together ideas from behavior-based robotics, control theory and machine learning and combines them with models from ethology, psychology and neurobiology in an attempt to synthesize a complete, artificial nervous system for a simulated artificial creature. It is argued that an intelligent system cannot be based on a single general principle, but requires a large set of interacting systems. The main goal of the thesis is to identify these functional subsystems and to develop computational miniature models of them that can be combined into a complete system. It is shown how goal-directed behavior can be categorized as appetitive, aversive, exploratory or neutral. This classification is a step away from a single hedonic dimension, and gives a richer framework for understanding reactive behavior. A number of learning mechanisms are developed that take this new framework into account, and it is shown how these mechanisms can account for a large range of classical and instrumental conditioning experiments, as well as more cognitive processes such as category learning, exploratory behavior and cognitive mapping. The role of expectations in learning is emphasized to map out the way for more cognitive abilities such as planning and problem solving. It is also shown how categorical, procedural and expectancy learning can all be based on different types of matching between the actual and the expected sensory state. The central role of motivation and emotion within a cognitive theory is discussed, and it is shown that a central motivational system is necessary to coordinate behavior. (Less)