Modeling Instinct Using Expert Systems

Given that instincts are bodies of expert knowledge then the methods of expert system building should be useful in modeling them. In fact trying to build an expert system model of an instinct might be the ideal way to come to grips with the deep research problem of how instincts work. Modeling instincts is a research program.

A bird building a nest or a beaver building a dam are clearly cases of the application of expert knowledge. If anyone doubts this they should try building a bird nest. So are a horse grazing in motion, a crow cawing, a herbivore selecting plants to eat, or a woodpecker picking a drum tree, as I see it. Each of these cases is discussed in prior articles here. There are many other cases that could be mentioned. In fact defining the specific areas of instinctive expertise in each kind of critter is a research program all on its own.

Building an expert system primarily means specifying a set of rules that embody a specific body of expert knowledge. That this can be done, that such expert rules exist, is itself a great discovery. Often the individual rules are relatively simple. It is the combining of these simple rules that yields the complex behavior of expertise.

Normally the expert rules are found using a process called knowledge engineering. It involves a combination of interviewing experts and reading technical documents, such as manuals, handbooks and textbooks.

When it comes to horses or critters in general, this method of knowledge engineering is not available. Rather the approach has to be to ask what expert rules explain the observed behavior. This is likely to be significantly more difficult than simply interviewing an expert, but there is no reason it cannot work. Even asking the question is useful because it creates a systematic approach to understanding an instinct. It breaks the problem down into one of finding a set of simple rules.

For example, here are two simple rules that might help explain how herbivores decide which plants to eat:

Rule 1: If it tastes good it is probably okay to eat it.

Rule 2: If it tastes bad it is probably not okay to eat it.

Note that the use of "it" here will require being able to tell one kind of plant from another. This will require a set of rules of its own. The use of "probably" means that other rules might modify these rules, such that good tasting plants are not eaten and bad tasting plants are eaten.

The point is that instinct is often an innate form of expertise. Trying to find the simple expert rules underlying important instincts is a feasible research program. This will greatly increase our understanding of animal behavior. In the case of horses this understanding will facilitate the training and management of the critters.

What plant eaters eat and how to think about it

We recently observed some profoundly interesting eating behavior in groundhogs. A juvenile groundhog and its mother were grazing together. The mother started eating a plant and the juvenile came over and ate some of it out of her mouth. It then went and ate a great deal of the same plant, which we had not seen it do before.

It is possible that this was all a coincidence, but it looks like the juvenile learned that this kind of plant was edible from its mother. This raises the issue of how herbivores, including horses, know what to eat? It also raises the issue of how animals know how to learn?

The eating issue is interesting because the number of different species of plants is enormous, so instinct alone cannot say which are edible and which are not. Instinct can provide general guidance, beyond which there must be some some sort of learning process. but the learning process itself must be at least partially instinctive.

Here I am reminded of Chomsky's theory of language learning in humans. He argues that infant humans learn language far to quickly for the process to be one of inductive inference, that is by generalizing broad rules from narrow instances. There are too many different possible languages that fit the infant's limited experiences.

This also rules out trial and error learning, which we do not observe. Trial and error may occur for specific words, but not for learning the language as a whole.

Chomsky therefore concludes that all human language has an underlying structure that is known instinctively. If so then the vast array of different human languages are merely local variants on this universal underlying structure. Thus the child is not learning language per se, rather just the local variant.

The same may also be true when horses, groundhogs and rabbits learn what to eat. The basic framework knowledge must be instinctive, supplemented by a learning process that is also grounded in instinct. This is just two examples of an instinct being a body of basic knowledge. The challenges are (1) how to figure out what that knowledge is and (2) how to express it using human concepts and language, which may be very different from the critter's concepts.