Sian<<s system (1990a, 1991b) is involved in both individual and distributed learning. Each system learns individually, but if certain conditions arise interaction is initiated with other agents. The interaction is established via an interaction board, which plays a similar as in blackboard architecture systems. Here the agents can, for example, propose a hypothesis to the the interaction board.
Communication between the learning agents is whenever one of the agents has obtained a hypothesis and has sufficient confidence in it. This is considered as a good candidate to put to test. Opinions of the other agents are solicited with the objective of establishing a consensus. The rules can remain as they are, or they can be modified, or they can be withdrawn. The rules that have been agreed upon represent a consensus of the group and appear in the "integrated theory".
This work differs form the other two presented earlier in various aspects. First, the author has elaborated an interface through which the individual agents communicate. Introduction and retraction of hypotheses to/from the interaction board is achieved using the operators
PROPOSE, ASSERT, WITHDRAW, ACCEPT
Evaluation of hypotheses is accomplished using the operators
CONFIRM, DISAGREE, MODIFY, NOOPINION,
while AGREED modifies a state. Each hypothesis is characterized by a NET-VALUE calculated on the basis of the opinions of different agents (CONFIRM, DISAGREE, MODIFY, NOOPINION) and the confidence values associated with each operator.
We notice that all three systems discussed in this section (i.e. Gams<<s, Brazdil & Torgo<<s and Sian<<s) use some particular method for assessing the usefulness of a given rule on the basis of evidence presented by different agents. Further work could be done to present a more detailed comparative study.
As we have mentioned earlier Sian<<s system differs from Brazdil and Torgo<<s in one important aspect. The agents are allowed to interact in the learning phase. This seems to make sense, particularly if we are interested to save some agents<< effort associated with learning. The earlier a potentially good hypothesis is put to test and possibly accepted, the better.
When considering testing in a multi-agent environment, it is necessary to distinguish between centralized testing (done by one agent) and distributed testing. Testing against all data available does not necessarily imply a centralized solution. A particular hypothesis can be sent to different agents. Each can then update the information received. A global view can be thus built up by passing a hypothesis from one agent to another.
In Sian<<s system each agent tests the proposed rule against his own data. A global view of each rule is then built up on the basis of a number of local views. A question arises whether this built-up view is equivalent to the global view that could be obtained by centralized testing. Brazdil and Torgo<<s system seems to satisfy this criterion. Each agent could update the qualitative and quantitative characterixation of the given rule and then pass this information to the next agent. This information is the same as the one generated during centralized testing.
Further work could be done here. A study of cost-effectiveness of the two methods could be made, taking into account:
- the effort of transferring the local views / instances to one agent,
- the effort of evaluating a given hypothesis (using instances / local views),
- net increase of confidence for some particular method.