Biological Neural Networks in Invertebrate Neuroethology and Robotics Editors: Randall D. Beer, Case Western Reserve University, Cleveland, OH Roy E. Ritzmann, Case Western Reserve University, Cleveland, OH Thomas McKenna, Office of Naval Research, Arlington, V A A c a d e m i c Press, Inc., Harcourt Brace Jovanovich, 1993 Reviewed by: Glenn Becker Thomson Technology Services Group 1375 Piccard Drive #250, Rockville, M D 20850 gbecker@thomtech.com This book is the result of a workshop held in the fall of 1991 at the National Academy of Sciences Study Center in Woods Hole, MA. The workshop was entitled "Locomotion Control of Legged Invertebrates" and it was sponsored by the Office of Naval Research. The purpose of the workshop was to bnng together researchers studying neural control of sensory and motor systems in invertebrates and researchers building legged robots. This type of cross-discipline work is often beneficial for both groups. Researchers in robotics and artificial intelligence borrow regularly from biological systems. Likewise, researchers studying neurological systems often benefit from computer models of control systems. These models can be controlled and manipulated much more easily than their biological counterparts. The book is divided into four distinct units. Each unit has four to five chapters. The first unit focuses on the neural control of leg movement. The reader of Unit 1 should already have a good understanding of robotics and control systems. Knowledge of how neurons and simple neural networks work is also useful but the authors explain these concepts for those who do not have this background. Much of Unit 1 deals with control systems for static and dynamic structures in biological systems. Simple neural networks from invertebrates like roaches, lobsters, and leeches are used for most of the biological examples throughout the book. Unit 1 is a good overview of the current understanding of biological control systems and contains plenty of references to other materials. Those who are looking for mathematical models and data to support those model will be disappointed, however. The book in general does not show a lot of raw data or mathematical equations. The second unit looks at the neural control of orientation and simple reflex network operation. This unit requires a reader with a good background in biological neural networks. The chapters in this unit focus on how biological control systems determine the orientation of external stimuli and the orientation of limbs relative to the body and other limbs. Controlling limb orientation also leads to a discussion on low-level goal seeking. Although interesting, most of the chapters in Unit 2 look at neural networks and show how they coordinate movements but they do not tie this to locomotion. The third unit focuses on using computers to model biological neurons and simple networks of neurons used for building control systems. The authors show the network configurations and neuron behavior that they use in their models. One of the chap- ters shows how a digital neural network can also be used as a real-time control system that is adaptable. The neural network can be tuned to operate better by changing the connection weights among the neurons. One of the main themes through this unit is how frequency modulated (FM) biological neurons can be timed and synchronized in a network to control leg and body movements. This timing and synchronization among neurons is what enables them to control complex real-time activities like walking. As mentioned before, the authors do not supply enough raw data for their models to allow the reader to reconstruct their experiments. The fourth unit is a review of legged robotics and some current projects using biologically inspired control systems for robotic control. The review of legged robotics in this unit is definitely worth reading for those new to the field. The authors put the evolution of the technology and the wide variety of techniques in perspective. This last unit gets away from the biological neural network discussions found in the first three units and focuses on more on traditional robotics and control systems. This book is an excellent reference for any researchers or students interested in insect or robotic locomotion. It offers sufficient explanations of electronic control systems and biological networks of neurons to bridge the gap between these two fields. It highlights the areas where these fields can successfully borrow from each other. The work presented in this book could also be used for building control systems other than locomotion. Similarities between eye movement and camera control and hand movement and articulator control have also been similarly exploited to improve robotic systems. Ongoing work in this area includes finding transforms or conversion techniques to move among biological neural networks, digital neural networks, and traditional control systems. Each of these three systems can be used to control real-time systems. Biological neural networks are used by animals to control their movements. Traditional feedback control systems are used widely for controlling robotic systems. Digital neural networks seem to fit in between these two paradigms, being used to simulate biological neural networks and as feedback control systems with the ability to adapt over time. This book shows the potential benefit of a set of transforms that could take a biological neural network and convert it into a digital neural network. The digital neural network could then be trained and customized for a particular task. This digital neural network could then be converted into a traditional feedback control system. Control systems can be analyzed for stability which is either difficult or impossible to do with digital neural networks. This stability makes feedback control systems popular for use in most robotic systems. Adaptive control systems are required in dynamic environments where the robot must be able to learn how to avoid new obstacles and remember where they are and how to avoid them in the future. These adaptive control systems can be implemented as digital neural networks which are closer to the biological neural networks that allow animals to deal with their dynamic environments successfully. S I G A R T Bulletin, Vol. 7, No.4

/lp/association-for-computing-machinery/biological-neural-networks-in-invertebrate-neuroethology-and-robotics-neTPkg04Bk