[1] Discussions about this topic, and many others, where held at the Ponte de Lima Workshop on the Foundations of Expert Systems [ESF]. The author acknowledges an intellectual debt to the participants in the Workshop, and specially Doug Lenat and B. Chandrasekaran regarding the topic of this report. This section, particularly, can be seen as a continuation of the discussions in the Workshop. The conclusions and proposals in this report, however, depart from the Workshop's participants opinions and are the author's responsibility.

[2] Essentially a MOP is like a script, an stereotyped sequence, but its components now may be shared among a number of MOPs. The basic components, scenes, describe events that take place in a single location, with a single purpose, and in a single time interval [SCHANK89].

[3] The use of frame-based systems in AI has two different aspects. Frames have been intensively used in AI research, while in AI applications (mainly expert systems shells and tools) they have a lesser success. This may be due to the fact that frames are not used in a principled way as a knowledge representation level but are used as they were providing a mere object-oriented programming environment [FIKES]. This problem is an instance of the general issue of representation languages: they do not offer the expressive power of a general purpose programming language. As a result programmers hack an escape to the underlying (general-purpose) language [NOGOOD]. This problem has also sometimes been defined as if the representation designers were not able to anticipate what is important for the users, but that is a wrong approach. The central issue is that representation languages do not offer what programming languages offer, and hybridizing both is the common practice. Of course, representation languages is also a major topic for MMP (see [[section]]6).

[4] We will make further on a distinction between deliberate and spontaneous inference. While still related to this deliberate/memory contraposition (see [[section]]6) it deals with different issues and stems basically from discovery systems like [LENAT] and [HAASE]. On the other hand, the MMP approach supports the vindication of memory-based architectures pointed out by Chandrasekaran.

[5] Other related projects, with a more limited scope, are the Carnegie-Mellon University World Modelling Project, and the Stanford University Knowledge Systems Lab Engineering Design Project.

[1]Both clichés are a subtype of FILTER, which is a cliché using a POSER cliche that imposes a partial order, and a SELECTOR cliché that selects the maximal of that partial order. The partial orders used in the example are those created by the observation cost and the informativeness likelihood criteria.

[*] May be it is worthwhile to stress again that the experinece based paradigm is an AI paradigm, not a new paradigm for machine learning, i. e. the constitutive elements and distinction posited by the experience-based paradigm and the MMP approach are to be assessed in the context of intelligent behavior and not in the more restrictive context of learning situations.

[*] A possible definition for analogical reasoning is that the transfer of experience is from a domain D to a new and different domain D' while case-based reasoning transfer is from a case C to a Case C' both in the same domain D. The problem with a term like analogy is that it is overloaded with different meanings and no definition suffices. Research in analogical reasoning has to uncover the different analogical methods and characterize them.

[6] This issue is raised again, inside a different perspective, in [[section]]6.2. The need for deliberate inference when no spontaneous inference can automatically cope with a situation, brings to mind SOAR's need to reflect on its problem-solving when an impasse is reached. I think studying this analogy is a good research issue.

[7] Where essentially means that maybe also deductive inference plays some minor role on spontaneous inference, or that some sequentiality is found in the spontaneous inference, and vice versa for deliberate inference.

[*] Also, it can be seen, in relation to the Society of Mind, as the process for recovering old partial states of the mind

[8] The computer metaphor of compilation can be used to characterize spontaneous inference as compiled knowledge. I think this is confusing, since compilation is a simplifying translation, while spontaneous inference runs quickly merely because it needs not to access or create intermediate structures created on the fly (they are already there) [ARLO]. For another detailed discussion on the metaphor of knowledge compilation see [ESF].

[*] The learning by accumulation postulated by the Society of Mind [MINSKY] and

[*] The spontaneous inference may achieve the goal only partially, as in classification by partial matching. From that, a deliberate inference may try to solve the mismatching taking this as its goal.