Issues

• Will compilation occur?

  Decision theory, although powerful and precise, often leads to inefficient, if not intractable, mechanisms. The inclusion of a condition-action execution architecture provides a highly reactive component, but it remains to be seen how much knowledge of this type the agent will actually possess. In general, most of the a priori knowledge will be of the more uncompiled types. If the system could learn the more compiled forms, then clearly its knowledge will shift toward a more efficient representation. However, there is no current implementation of this compilation process, and it remains to be seen how effective it can be.

• Will the behavior be adequate and coherent?

  The agent's behavior depends greatly on the choice of utility functions. Since the execution architectures make local decisions, it is possible that in the long run, goals will remain unsatisfied (this myopia is mentioned by the authors). For instance, in the current state, a certain action is preferrable since it advances toward goal A, although it moves away from goal B. In the resulting state, perhaps goal B actions have higher utility, and reverse the previous results. In addition, there is also the problem of local minima. Of course, correct choice of utility function can correct both problems, but sometimes the correct choice is unkown.

• Should the EAs share knowledge?

  In general, the execution architectures use only a subset of the different types of knowledge. When considering only one atomic action, this restricted access would seem to be acceptable. However, when the planning window stretches over action sequences, they should perhaps combine different types. For instance, a condition-action rule may provide a good (and fast) action choice from the current state, but there may be no such good rules for the resulting state. On the other hand, Decision-Theoretic choices for the current state could require a great deal of computation for the first choice, but could provide the next step satisfactorily. It would seem that the desire for a combination of the two planning steps would override the wish for modular organization.