Reviews Intelligence as Adaptive Behaviour: An Experiment in Computational Neuroethology By Randall D. Beer Academic Press, 1990. $29.95. Reviewed by David M. W. Powers DFKI University of Kaiserslautern W-6750 KAISERSLAUTERN, GERMANY powers@dfki.uni-kl.de, Moving on to the biological, and in particular neuroethological, background, Beer does an excellent job of laying the groundwork for the research he presents. His carefully defined heterogenous locomotion controller reproduces surprisingly well the different gait patterns observed in nature with the help of excitatory and inhibitory interplay of of backward and forward angle sensors, and represent a significant result. However, the feeding controller is not so parsimonious and does not represent any real advance. Notwithstanding, it does display some points of interest. The adaption that lends its name to the book comes in here. The cockroach has an arsenal of different behaviours, and the controller selects between them on the basis of external and internal sensor/state information. The cockroach is able to find food and negotiate obstacles, and thus survive in a dynamically changing environment. A second point where adaption has been demonstrated is in various "lesion" studies where parts of the locomotion system have been disabled. For example, the cockroach has been able to continue to move with only a slightly altered gait when the forward angle sensors have been disabled. On the other hand, motor lesions affecting a single leg (equivalent to "amputation") was far more devastating than it would be for a real cockroach. In this case we are looking at survival in the face of changes to the organism's own systems: robustness. Let us come back to the title of the book: "Intelligence as Adaptive Behaviour." How intelligent is the behaviour described here? Robustness is not necessarily intelligence--consider asixcylinder car chugging along on five cylinders and the six-legged insect loping along on five legs. The robustness is an inherent outworking of the design. What about an insect that heads towards the smell of food until it encounters an obstacle, and then follows the outline of the obstacle until it is able to proceed in the direction of the food once more? Indeed there is a measure of intelligence here, or apparent intelligence. It this intelligence which Beer ascribes to adaptive behaviour. But it must be made clear that it is rule-based behaviour and does not involve learning. We've struck once more the ubiquitous problem of Artificial Intelligence: once we understand the rules which lead to a certain apparently intelligent behaviour, we are much more unlikely to call it AI. Intelligence as Adaptive Behaviour."An Experiment in Computational Neuroethology is based on a Case Western Reserve Ph.D. dissertation. The two parts of the title reflect the two aims of the author's argument and research: "a view of intelligence which is some somewhat different from the traditional one" and "a particular methodology for the construction of autonomous agents." Whilst in relation to the latter goal we are presented with useful insights into the successful construction of P. computatrix, a simulated cockroach in a simple graphical world, the former is not met at all well. One is led to suspect it is an afterthought, an outgrowth of the traditional literature review, or even just an exercise in marketing and packaging. The difficult part of writing this review is that I agree largely with Beer's position. Furthermore, Chapter One, where the adaptive nature of intelligence is addressed raises a good many important issues through quotes from many sources. What is somewhat frustrating is that he doesn't really attempt to clarify his own position until Section 1.4, and even there he depends on secondhand quotes. In summary, Beer takes the position that intelligence goes far beyond the traditional AI of Chess Programs or Theorem Proving, and that we have been largely negligent in ignoring the intelligence displayed by subhuman creatures, in this case the cockroach. What characterizes intelligence on this view? Beer's answer is "adaptive behaviour," the ability to deal with changing and unforeseen circumstances, a corollary is that behaviour is "situated" and concerns interactions between the agent and its environment. Interestingly, for the sort of adaptive behaviour he considers, a "structural congruence" between the internal and the external dynamics is, he claims, sufficient. This clearly contrasts with the idea that adaption has something to do with learning, although his primary discussion of learning vis-a-vis adaption is buried in the middle of his description of his simulation (pp. 61-2). Here he takes issue with the "overemphasis" of learning in neural network research and justifies his use of "heterogenous neural networks" in contrast with the usual "homogenous" variety: "Only once the proper neural architectures for controlling the behavior of autonomous agents have been uncovered can we begin to examine the ways in which the selective introduction of plasticity will increase the flexibility of the resulting controllers." (Interestingly, there are whole literatures on Symbol Grounding and Modularity which Beer does not touch.) Principle-Based Parsing: Computation and Psycholinguistics Edited by Robert C. Berwick, Steven P. Abney, and Carol Tenny Kluwer Academic Publishers Boston/Dordrecht/London, 1991. Pp. vii + 408. Reviewed by Michael A. Covington Artificial Intelligence Programs The University of Georgia Athens, Georgia 30602 1. Currently:Visiting Fellow, ITK, Tilburg University, NL-5000 TILBURG, HOLLAND, powors@kub.nl mcovingt@uga.cc.uga.odu S I G A R T Bulletin, Vol. 3, No. 2

/lp/association-for-computing-machinery/book-review-intelligence-as-adaptive-behaviour-an-experiment-in-Y6poFmfeYE