Edited by Francisco Varela and Paul Bourgine
The MIT Press, Cambridge, Massachusetts and London, England 1992, xvii + 515 pp., œ49.50
This volume contains the papers presented at ECAL 91 - the first European meeting on artificial life which took place in Paris in December 1991.
Artificial life is presented by its proponents as a new discipline that emerged from the pioneering work done at the Santa Fe Institute, USA. However, as the editors of this volume point out in their introduction, the field has its roots in cybernetics in the 1950s with the work of people such as Ross Ashby and Grey Walter, and it has remained a strong theme of cybernetics ever since. The link between artificial life (AL) research and artificial intelligence (AI) has always been strong with AI people tending to concentrate on intellectual intelligence and the AL people more on sensorimotor intelligence.
Varela and Bouillon argue that AL would be served best as a research effort if it were seen as concerned with autonomous systems rather than purely with issues of artificial life as such. The theoretical focus they offer involves investigating the modes of functioning that underlie viability and autonomy and they suggest that a dynamical systems approach is the way forward. This view links with Varela's earlier work with Maturana on autopolesis and operationally closed systems in which autonomy is seen necessarily to involve selfreference, self-production and the development of self-identity. This view is in strong contrast with most work in traditional AI which sees intelligence as something that can be engineered and fed into systems without any need for the system itself to define and create things like its own meanings and knowledge. Increasingly, however, there are people in AI who appreciate and support this different approach - Terry Winograd is perhaps a prime example. In my opinion AI has a lot to learn from AL research and several of the papers in this volume offer refreshing and stimulating approaches that I feel many AI workers would do well to study.
The book contains 57 papers, including the editors' introduction. The papers are arranged in five sections: Autonomous Robots, Swarm Intelligence, Learning and Evolution, Adaptive and Evolutionary Mechanisms, and Epistemological Issues and Conceptual Foundations. Within the space available in a short review it is not possible to discuss all the papers, so I will mention just a few that caught my attention.
I liked Pattie Maes' approach discussed in her paper "Learning Behaviour Networks from Experience". She presents a model for motivational competition between and selection of behaviours in an artificial creature. This model uses a behaviour network which is built up through learning and which encodes information about predecessor successor and conflictory links among possible chunks of behaviour in such a way that the most relevant behaviour sequence for a particular task and environment emerges from the network. The model implements behaviour selection in a completely distributed way through competition among behaviours for activation energy. As she says, this is a more flexible, faster and fault-tolerant approach than those of traditional AI.
Eric Dedieu and Emmanuel Mazer in their paper "An Approach to Sensorimotor Relevance" raise the important issue of how to make a robot's perception and knowledge belong to it rather than to the programmer's view of the task and domain. To illustrate the problem, they relate an incident in which a robot programmed for light-seeking unexpectedly also exhibited object avoidance behaviour. This was later explained by realizing that the lighting used was such as to cast shadows around objects so leading to object avoidance as a side-effect of darkness avoidance. They emphasise the need to treat perception and action as interdependent and contextually defined and always to be seen from the robot's perspective and not ours
The section on learning and evolution has papers with a heavy emphasis on genetic algorithms. In their paper "The Royal Road for Genetic Algorithms: Fitness Landscapes and GA Performance".
Melanie Mitchell, Stephanie Forrest, and John Holland address the key issue of trying to improve our understanding of why genetic algorithms perform as they do and what makes a GA application successful. What they outline is a principled research strategy for systematically investigating ' fitness landscapes" and how well Various genetic algorithms cope with them in order to develop a system of sound GA design rules and methods.
Jakob Skipper reports interesting work in his paper "The Computer Zoo - Evolution in a Box". He describes a computerised ecosystem in which programs written in a special language, which is more robust under mutation than normal programming languages, interact as in an ecosystem. His experiments show symbiotic and parasitic relation ships, the self-assembly of flocks of programs, punctuated equilibria and some rudimentary predatory behaviour
In the section on conceptual issues I liked the thoughtful papers by Robert Davidge - "Looking at Life", and Claus Emmeche - "Life as an Abstract Phenomenon: Is Artificial Life Possible?". Davidge makes a number of good points - for example: AL needs to attend to its conceptual and metaphysical foundations; the environment of the naturally living is mainly other living things and living things alter and enrich their environments, features which are generally lacking from most AL work; and that the physics that underlies computational processes and those of naturally living things are almost at opposite ends of a spectrum of mutability which may raise important restrictions for computer based AL. in his paper, Emmeche raises the fundamental question: Is life a property of the material structure of a living system or an abstract form of organization that can be realized in media both organic and non-organic? The unmentioned assumption of most AL researchers seems to be that life is an abstract organization - Emmeche raises some interesting points to question this view.
This volume is a "must" for anyone with an interest in artificial life research. r think it also has a much wider appeal and I feel that many people concerned with the future of AI and information science in its broad sense will benefit from it. Of course, those from the biological sciences who have an interest in the conceptual and philosophical basis of their discipline will also find much in this book to challenge and stimulate them.
Michael Elstob,
Brunel University,
January 1993.