Classical Artificial Intelligence failed to build systems which could scale up to meet real world problems. Today similar problems face researchers using alternative approaches such as Neural Networks or Evolutionary Computing. A key theme running through my research is the attempt to model the ways in which nature has evolved complex intelligence.
In models based on Neural Networks or Genetic Algorithms the researcher defines a task or a fitness formula and applies a "training" algorithm which produces networks or digital organisms satisfying the researcher's requirements. There are however good reasons to believe that this approach, which can be extremely effective for simple problems, is unable to produce "organisms" with complex capabilities. Complex biological organisms acquire their abilities step by step. In my research I have attempted to simulate this process using hybrid GA/neural network models. This research suggested that the task of designing a sequence of training or evolutionary procedures to achieve an arbitrarily complex goal may itself be intractable.
Biologists often work on the tacit assumption that natural organisms passively adapt to the requirements of their environment. Waddington and other authors have suggested, on the other hand, that populations of organisms can play an active role in "selecting" the niches they inhabit. In a series of simulations I have shown how niche selection can facilitate artificial evolution. This work suggests that evolution may be less powerful than Darwinians (and computer modelers) sometimes lead us to believe. Biological evolution only solves problems which are in some way pre-adapted to the capabilities of a population. Maybe it is not reasonable to expect computer models to do better.
Much work in Evolutionary Computing is conducted using a static environment. This is biologically unrealistic. All lineages of organism existing today have evolved in a sequence of different environments. My research suggests that in a system allowing niche selection, complexity will only emerge at a critical rate of environmental change. At rates below the critical threshold populations concentrate in very simple niches; at very high rates of change organisms are unable to adapt; at rates of change close to the critical level, the majority of organisms continue to occupy simple niches while a minority develop complex structures and capabilities.
Complex biological organisms have highly redundant genomes. There is, what is more, no clear relationship between the complexity of the organism and the length of the genome. Long genomes may be due to inefficient evolution or "genetic parasitism". It has also been suggested, however, that genetic redundancy can contribute to evolvability. We are currently investigating this hypothesis by manipulating genetic redundancy in populations of digital organisms and measuring the effects on performance.
A-Life models frequently produce "interesting results". Often, however, the models are far-removed from the specific scientific problems faced by biologists and psychologists. In a series of experiments we are using A-Life techniques to evolve organisms with the ability to perform classical tasks in animal psychology. To date we have studied simple cases of foraging and detour behavior. In both cases we have generated organisms which precisely replicate the behaviors observed in psychology experiments while using neural structures which are substantially simpler than those posited by many psychologists.
Much of my research has been based on Artificial Neural Networks with a simple feed-forward architectures. This kind of architecture is biologically unrealistic (all biological neural networks contain significant numbers of feedback circuits) and can only simulate simple stimulus-response behaviors. We are currently designing a research program to investigate "recurrent networks" with a "small world architecture" (many local and a small number of long distance connections). In this work, we intend to simulate a number of tasks of interest to psychologists, in particular in the field of "prospective memory" (the ability to remember a task to be performed at a future time).
Antonio Damasio has recently emphasized the essential role of proprioception in animal and human intelligence. In most A-Life models, on the other hand, organisms are provided with sensors reporting information on the outside world but have no knowledge of their own body or behavior. One well-studied example of the use of internal knowledge in the regulation of behavior is weight regulation in higher animals. The neural and molecular mechanisms underlying this self-regulation are relatively well-known. We are currently in the early stages of research aimed at evolving artificial organisms with the same ability. This work, it is hoped, will provide insight into the role of proprioception in the evolution of intelligent behavior.