One problem that persists in the design of systems that are not only knowledge-intensive but also must support multiple domains, is the choice of a suitable knowledge representation format. The problem lies in many fronts:

Different types of knowledge require different types of representation. This is addressed by hybrid representation systems [Heylighen 91]. [Minsky 81] [Winograd 75] [Woods 75];

Different types of knowledge require different kinds of reasoning. This is addressed by the use of multiple inference engines, and intelligent "dispatching" systems [Carroll 87] [Gleiz 90];

Knowledge acquisition and maintenance modules of the system are usually so hard-coded to a specific application (with pre-defined knowledge and knowledge types) that sustainability of the system is put in question. This is addressed with intelligent user interfaces [Ferraz 89] [Rissland 84];

Knowledge management usually implies the "internalization" of knowledge and data files, that is, any bit of information must be reformatted, re-classified and some times stored for private use of the system, creating a high impedance between the system and the outside world that further limits sustainability. This is addressed by non-obtrusive metadata strategies [Davis 77] [Ferraz 92].

Regarding knowledge representation paradigms, my approach is to test a set of criteria (adjusted by trial and error) that will build a library of default representation formats for each new "knowledge unit", in the domain of impact assessment considered by the system. For instance, knowledge about primary and secondary consequences of infra-structure shortfalls and of each alternative action, is more about causal relationships (if truck traffic and weak pavement than new road is needed) than about knowledge in depth about entities or objects (roads, trucks); this points towards a rule-based representation and reasoning. Other knowledge domains may depend on much weaker cause-effect relationships and be instead more based on precedent experience (like border cases in environmental law applications), pointing towards a case-based representation and reasoning. Yet other domains may be based on in-depth knowledge about entities, or objects (like land uses, or parametric description of water treatment systems), hence pointing towards the use of object-oriented or frame-based representation and reasoning [Booch 91].

In Table 1, I present a summary of the different knowledge representation models, the kind of inference (reasoning) engine usually associated with each, and the more suitable system dynamic context.

Table 1. Models of Knowledge Representation

Reflecting the earlier "general problem solving" orientation that prevailed within artificial Intelligence, many authors favor this or that model of representation as the most promising for any domain. The discussion concerning the relationship between representation and the world of applications is still going on [Pearce 92] [Aiken 91] [Davenport 91] [Gleizes 90] [Jaffe 89], and it remains as an open question. To build such a library of links between domain and representation, one needs to associate with each knowledge unit a descriptor about itself, or "metaknowledge" descriptor [Davis 77]. For the sake of tradition, I will use the term metadata in the wider definition that will include metaknowledge.