As has been suggested in the previous sections, communication plays rather an important role in multi-agent learning systems. It may supply the agent with valuable information and thus avoid "re-discovering the wheel". Communication need not, however, bring about benefits. It is thus important to study this topic in its own right. Although this topic exceeds the objective of this paper, we would like to make several observations here.
It is important to distinguish between the issues related to form of the language used between agents and the actual statements in that language. Here we make a similar distinction as when talking about natural language. There is a difference between problems related to structure of English and particular piece of text.
The issues related to the language itself could be viewed as issues of interfaces between agents. It is necessary to decide what kind of statements the agents should be able to generate and comprehend. For example, one could decide that the operator PROPOSE(H,C) should have a certain meaning. In Sian<<s system this operator adds hypothesis H (and the associated confidence C) to the interaction board.
Obviously, the design of interfaces is closely related to the design of the architecture of the whole system. The operator PROPOSE plays a specific role in the system for which it was designed. A question arises whether some set of basic communication primitives could be found that would be generally useful in multi-agent learning. This would have the advantage that it would make it easier to compare different approaches. Of course, one could always add extra primitives, or define other constructs in terms of the existing core primitives, if this was required in some specific system.
The second kind of issues are related to the problems of interpretation and meaning of agent<<s statements. As Shaw and Gaines (1989) have pointed out, same term can have different meaning for different agents. This situation is called a conflict. Different terms may, however, have similar meaning. This situation is called a correspondence.
Work of Brazdil and Muggleton (1991) is concerned with the problem of resolving certain language differences between agents. The agents are not only presented with different situations from which they can learn, but also, employ a somewhat different terms in their description of the (simulated) world. For example, one agent uses the predicate father(..) while the other parent(..). If we use Shaw and Gaines<<s terminology, there is a problem of correspondence. Brazdil and Muggleton show how these language differences can be overcome. It is shown that standard machine learning techniques can be used to acquire the meaning of undefined concepts.
There are interesting relationships between inductive learning and communication. There interplay mentioned here is of a different kind than the one discussed in Section 3. There we have discussed different ways communication can supplement learning. Here we are concerned with the possibility of resolving certain problems of communication using learning.