Recent years have seen a proliferation of intelligent agent applications: from robots for space exploration to software agents for information filtering and electronic commerce on the Internet. Although the scope of these agent applications have blossomed tremendously since the advent of compact, affordable computing (and the recent emergence of the World Wide Web), the design of such agents for specific applications remains a daunting engineering problem;Rather than approach the design of artificial agents from a purely engineering standpoint, this dissertation views animals as biological agents, and considers artificial analogs of biological structures and processes in the design of effective agent behaviors. In particular, it explores behaviors generated by artificial neural structures appropriately shaped by the processes of evolution and spatial learning;The first part of this dissertation deals with the evolution of artificial neural controllers for a box-pushing robot task. We show that evolution discovers high fitness structures using little domain-specific knowledge, even in feedback-impoverished environments. Through a careful analysis of the evolved designs we also show how evolution exploits the environmental constraints and properties to produce designs of superior adaptive value. By modifying the task constraints in controlled ways, we also show the ability of evolution to quickly adapt to these changes and exploit them to obtain significant performance gains. We also use evolution to design the sensory systems of the box-pushing robots, particularly the number, placement, and ranges of their sensors. We find that evolution automatically discards unnecessary sensors retaining only the ones that appear to significantly affect the performance of the robot. This optimization of design across multiple dimensions (performance, number of sensors, size of neural controller, etc.) is implicitly achieved by the evolutionary algorithm without any external pressure (e.g., penalty on the use of more sensors or neurocontroller units). When used in the design of robots with limited battery capacities , evolution produces energy-efficient robot designs that use minimal numbers of components and yet perform reasonably well. The performance as well as the complexity of robot designs increase when the robots have access to a spatial learning mechanism that allows them to learn, remember, and navigate to power sources in the environment;The second part of this dissertation develops a computational characterization of the hippocampal formation which is known to play a significant role in animal spatial learning. The model is based on neuroscientific and behavioral data, and learns place maps based on interactions of sensory and dead-reckoning information streams. Using an estimation mechanism known as Kalman filtering, the model explicitly deals with uncertainties in the two information streams, allowing the robot to effectively learn and localize even in the presence sensing and motion errors. Additionally, the model has mechanisms to handle perceptual aliasing problems (where multiple places in the environment appear sensorily identical), incrementally learn and integrate local place maps, and learn and remember multiple goal locations in the environment. We show a number of properties of this spatial learning model including computational replication of several behavioral experiments performed with rodents. Not only does this model make significant contributions to robot localization, but also offers a number of predictions and suggestions that can be validated (or refuted) through systematic neurobiological and behavioral experiments with animals.
展开▼
