An artificial insect

across the floor, you may be in no mood to marvel at the intelligence guiding the insect's behavior. Certainly that intelli? gence is quite unlike human cognitive ability; the insect does nothing that most people would recognize as ^think? ing/' From the point of view of some? one trying to create an artificial intelli? gence, however, the insect/ s performance is impressive. The insect perceives changes in its environment (such as a sudden increase in illumination); it chooses a response to those changes (fleeing for cover); and it puts the re? sponse into action (by running across the floor). Each of these activities entails a multitude of lower-level tasks; for ex? ample, nmning requires the insect to co? ordinate the motions of its six legs. An insect also exhibits a variety of internal? ly motivated behaviors, such as seeking food or a mate. All of these capabilities put the insect/ s intelligence well beyond that of the most sophisticated robot. There are two main perspectives on the study of such insect behavior. First, one can ask how the insect does it. By combining techniques from neuro science and ethology, neuroethologists have begun to address this question. They have identified a few neurons and neural circuits that appear to have a role in specific simple behaviors. An insect has hundreds of thousands of nerve cells, however, and most of those cells are difficult to identify and probe individually. Furthermore, the relationship between a particular be? havior and the activity of a specific set of neurons can be extremely subtle and context-dependent. Most behav? iors require a variety of actions to be coordinated throughout an animal's body and so may involve a significant fraction of its entire nervous system. The obstacles to a detailed under? standing of the neural basis of animal behavior are thus substantial. The second view of insect intelli? gence asks what we can learn from bi? ology about the design of robust and adaptive control systems. Many man made devices would benefit from just the kind of skills that simple animals possess. For example, a robot for au? tonomous exploration needs to negoti? ate varied terrain, to navigate through unfamiliar territory, and to adapt to contingencies unanticipated by its de? signers. Even though the computers available for such tasks are many times faster than the neurons of an animal nervous system, autonomous robots are far inferior to insects and other ani? mals in basic survival skills. The two perspectives are combined in the approach called computational neuroethology. This new approach concentrates on how simple nervous systems give rise to complex behaviors when an animal adapts to a changing environment. Behavioral and neurobi ological ideas gleaned from the study of relatively simple natural animals guide the construction of artificial ner? vous systems for controlling the be havior of autonomous agents. Such an autonomous agent demonstrates a ca? pacity to change its behavior, to adapt, as it interacts with its environment. Our own efforts in computational neuroethology have resulted in the cre? ation of a computer-simulated insect called Periplaneta computatrix?named after the American cockroach, Periplane? ta americana. The simulated insect "lives" in a simulated two-dimensional environment, visible on a computer screen. It has an artificial nervous sys? tem capable of performing such funda? mental tasks as walking and following the edge of a wall, so that the animal can explore its environment autono? mously. P. computatrix is also capable of feeding itself when it is low on energy. The mechanisms underlying the walk? ing and the feeding behaviors are based in part on specific neurobiological data. To the extent that some of our neural circuit designs remain faithful to the neurobiology that inspired them, the in? sights gained from the simulated cock? roach may help us to understand the workings of natural nervous systems.