Artificial Intelligence

~ ~ Artificial Intelligence (AI) is a term which covers the issues relating to the development of intelligent agents in this thesis. The purpose of these agents is to simulate members of the CE design team. There are two important issues which arise from this role. The agents must have particular roles and the agents must be able to communicate. The literature on these roles is covered in this Section.

Before discussing particular aspects of the intelligent agents in a design system, it is necessary to show the applicability of using expert systems as design aids. The development of the Computer-Aided Mechanical Expert System (CAMES) is discussed in [Campbell et al 1991]. The motivation for that project was their difficulty in finding experts to do machine design. This suggested an expert system to fill this gap. The first machine chosen as the design domain was one of the hardest for an expert designer to design. Campbell et al hoped to find all the problems this way and make development of further expert systems for less complicated machines easier. LISP was chosen as the AI language, and AutoCAD, which is written in LISP, was used as the design drawing system. The system maintains a library of components used, including spatial attributes. The expert system helps to design the machine. CAMES is a successful example of how expert systems can be used as design aids.
Brown [1992a] describes the task of design and the applicability of various AI techniques, including Knowledge Representation, Constraint Satisfaction, Search, Learning, Case-based Reasoning &Analogy, and Qualitative Reasoning, to the design task. Several roles for expert systems in a design system are identified in that paper. Some of these were adapted for use as the various intelligent agents in SNEAKERS. Each agent has a specific role, as described below:

1. Advisor - makes recommendations for the next decision based on the current state of the design. The user may ignore the advice.
2. Critic - compares the design to certain standards, and offers some criticism of the last action based on those standards.
3. Suggestor - takes the criticisms from the Critic and offers suggestions for satisfying them. Again, these suggestions may be ignored.
4. Analyst - offers numerical analysis to derive attributes, such as strength, cost, or size. These analyses are used when determining the success of the design.
5. Evaluator - evaluates the whole design from one perspective. Using the Analysts, evaluation determines how well the design takes into account the needs of that perspective.

In [Nii 1986a], the first part of this survey of blackboard systems, is a summary of what a blackboard system is and what its components are. The individual intelligent agents are referred to as Knowledge Sources (KS). The Blackboard data structure is a programming abstraction. It can be read by the individual KS's. KS's can also write to the Blackboard. The Control system is responsible for routing information to those KS's which can use it and decides which KS to use for a particular task. The actual implementation of these elements is different for different blackboard systems, but they are the essential ingredients. Nii also describes how this problem-solving method is applicable to many tasks, including the HEARSAY-I project, which was the first blackboard system developed. Its purpose was to recognize human speech.

[Nii 1986b] begins where the last article left off. It includes: HEARSAY-II, the successor to HEARSAY-I; HASP/SIAP, that maintained surveillance of surface ships and submarines from sonar data; CRYSALIS, which was designed to infer the three-dimensional structure of protein molecules; and TRICERO, which monitored an area of airspace for traffic. Nii then talks about skeletal systems, OPM in particular. She discusses applying Blackboards to the Scene Understanding problem, where the task is to label objects in a photo taken by a low flying aircraft. The final section deals with Knowledge Engineering issues which might lead to the decision to use a Blackboard approach. These include Problem Complexity and Ill-structured Problems. Nii suggests ways the approach could be used to formulate problems, as a system development tool, and as a research tool.
Blackboards are a part of work done by the Blackboard Technology Group. (See [Blackboard 1991].) As they state, blackboards allow separate knowledge systems to take information provided and either ask for information from others or give information to others. Their system is composed of a Controller and several Knowledge Sources (KS). The Controller uses its rules to determine which KS will run and when that KS will run. Each KS is also totally independent of the others. They can use different approaches to solving the same problem.
Durfee et al [1989] discuss the idea of Cooperative Distributed Problem Solving (CDPS), which is similar to the Blackboard problem solving approach. They discuss many issues concerning multiple distributed problem solving systems. These systems break a problem down into subproblems, in order to generate a solution to the problem. CDPS works through local problem decomposition. When a system is given a problem, it has the option of solving the problem it has been given by itself, or decomposing all or part of the problem and sending it off to another system. Contrary to this method, Blackboard systems work through global decomposition, with a Controller in charge of the whole problem. This is the main difference between CDPS and Blackboards. In CDPS, no system controls the whole process but each tells the others when to activate, whereas in a blackboard system, the scheduler has control.

Some of the important issues in the CDPS method are as follows: negotiation; functionally accurate cooperation; organizational structure; multi-agent planning; sophisticated local control; and formal frameworks. Negotiation is a major concern here. The system that has control of the problem sends a request specifying the problem to be solved, and those systems that have the resources to solve it respond, listing their capabilities, so the ``Controller''(for this individual subproblem) can decide among them.
Myers et al [1991] discusses the Intelligent Computer Aided Design System (ICADS) expert design advisor. This is a very extensive system which includes a user interface to a CAD system, six domain experts, a conflict resolver (i.e., negotiator), and a blackboard for handling communication. The user interface is a ``wrapper'' for a CAD drawing system, so that the user can interact with the design drawing and the expert advisors. The domain experts were written in CLIPS and each have a single machine dedicated to them. Therefore, they are truly distributed. The conflict resolver has the task of deciding among conflicting design decisions made by the domain experts. One of the major problems with the conflict resolver was giving it the ability to recognize when it had been in a certain situation before in order to avoid infinite loops of swapping values between conflicting experts. A blackboard was set up so that all of the experts would know where to voice their questions and answers. It is a separate system running on a separate machine.