The system described by Brazdil and Torgo (1990) attempts to integrate the knowledge acquired by individual agents. The integrated theory is then used by one of the agents to resolve problems.
The system works in three phases. In the first phase the agents go through individual learning. There are no interactions between the agents then. This phase is followed by knowledge integration. This process involves all agents in principle. Knowledge integration can be regarded as a special form of distributed (re-)learning. This process involves characterization of individual theories (or rules) on the basis of experimental tests. These provide the system with estimates of quality or utility of individual theories (rules). This method could be compared with the one used by Gams et al. mentioned earlier employing confidence factors. The quality estimates determine which theories (rules) should be included in the integrated theory.
Experimental results have shown that, in general, the integrated theory had a significantly better performance than the individual theories. We believe that this is due to the fact that redundant knowledge is properly exploited by this system. The knowledge integration method can be seen as a kind of "symbolic filter" for noisy knowledge (imperfect theories and noisy test data).
This approach differs from the one described earlier in several aspects. First, the system can resort to individual problem solving mode. Problems can be directed to the agent that has assembled the integrated theory (although this theory could be given to other agents too). Problem solving is thus simpler and hence the whole system is more "reactive" if we use the term from autonomous agent learning. It is not necessary to consult the whole community of agents before giving an answer. It is interesting to ask question why this should be so.
As we have mentioned earlier, different agents are called upon many times, but this is done at knowledge integration time. The result of knowledge integration is stored for future use. Consequently one need not consult different agents later. The system of Gams does not attempt to construct such a theory, and so it is necessary to solicit opinions of other agents at problem solving time.
There are arguments for and against each approach. The system described by Gams retains structured representation of knowledge. As individual agents update their knowledge, this immediately bears some effects on the opinion of the group. This is not true of the integrated theory. If one of the individual theories has been altered, the integrated theory may need to be revised. In a certain sense, the first approach has similar advantages as interpreting, while the second one has the advantages of compiling.
Sometimes it may be difficult or outright impossible to construct an integrated theory. Difficulties can arise particularly when the agents use different (and possibly incompatible) ways of representing knowledge.
Integrated theory represents a more compact representation of knowledge than the structured representation discussed earlier. Compact representations have obvious advantages. Simple theories are easier to communicate to other agents (including humans) than complex ones. They can also serve as a useful starting point in further learning.
Alternative theories are no doubt useful both in science and politics. Alternative theories often find their adepts, and it would be wrong to try to come up with one integrated theory that would explain everything. However, people would generally agree that there is a limit as to how many theories should be taken into account. Some theories may be just minor variants of others. In our view methods are needed that would determine whether some particular theory is worth keeping around as a useful alternative.