Chppter 4 : An Advisor Machine Intelligence and Human Control
4.3 Development Framework for an Advisor in Experimental Design
4.3.1 Identification
(i) Objective«
The identification step involves determining the problem domain and its decision-making process, characteristics, and suitability in principle to expert systems. It is also necessary to iden tify the knowledge engineer and the domain expert(s). These last two points may be resolved immediately. The knowledge engineer will be the author who will also be the initial expert. The
justification for this choice lies in the fact that the objective is only to develop the principle and framework for a support tool. If an acceptable framework is found then it should be possible to incorporate further knowledge ("expertise"). This will be addressed in chapter 5.
(ii) Domain analysis : simple intelligent reasoning model
The initial part of the identification process was addressed in the previous section when the appropriateness of an expert system framework to the domain of experimental design and analysis was considered. Within this domain, the context in which the decision-making process occurs is provided by the behaviour of an "expert" simulation analyst. In chapter 3, the following behaviour pattern was presented where the expert:
(1) - considers the model he is to experiment with;
(2) - defines the task/objective of his study;
(3) - designs an experiment to yield results that he expects will help his investigation;
(4) - executes his experiment, in the form of simulation runs, by carefully inputting the experiment configuration values (parameter values, starting and stopping conditions. number of replications, etc) and collecting results;
(5) - performs some preliminary analysis that may lead to possible replication;
(6) - analyses results;
(7) * presents conclusions.
This understanding of the experimentation process was acceptable when the consideration lay with the control of execution of experiments. However, since the focus is now on the intelli gent component, it is necessary to refine this model of the reasoning process. The concentration area will be steps (3),(4) and (5)-(6). Essentially there is intelligence involved before running a model that deals with designing the experiment, and there is intelligence involved after the run in the analysis stage. Therefore, conceptually steps (3),(4) and (5)-(6) can be characterised respec tively by :
- intelligence in design,
- control of execution, - intelligence in analysis.
The decision making process can be graphically modelled :
▼
runi^nd r..ulU collection refinement onolysi* of reeulte for elgniflconce
I ---
The outcome of this process is that results are always significant.
(ill) Domain analysis ; extended intelligent reasoning model
The previous model only addresses the specification of one experiment at a time. In gen eral, this pattern is part of a larger investigation that may require several experiments. Extending the previous model implies consideration of what happens before an experiment is specified and what happens after results have been collected.
? -
▼
(a) Intelligence in Design
This phase starts once the task has been defined by the user. His understanding of the model and task will lead him to consider a plan f< approaching this problem. This planning will be bom out of his simulation knowledge and experience and will be referred to as strategic rea soning. Strategic reasoning may provide at least two types of plans :
- to design a starting plan by specifying an experiment, in order to react to results;
- to design a global plan that specifies several experiments and then to analyse all results together.
The plans will be called strategies.
(b) Intelligence in Analysis
Different levels of analysis are required after an experiment is executed. - Checking for Significance : which may lead back to refinement
- Evaluation of Performance : once results have been collected, there is a need to analyse them and report their contribution. This may lead back to some more strategic reasoning.
- Analysis of the Investigation ; after all experimentation has been conducted, a more detailed analysis may be required.
(c) Model o f Intelligent reasoning
By combining the above, the following model may be derived as presented in figure 4.2.
This translates into the general framework of expert behaviour where the a n a ly st;
(1) - considers the model he is to experiment with;
(2) - defines task/objective of his study;
(3) - considers a strategy;
(4) - designs an experiment to yield results that he expects will help his investigation;
(5) - executes his experiment, in the form of simulation runs, by carefully inputting the experiment configuration values (parameter values, starting and stopping conditions, number of replications, etc) and collecting results;
(6) - performs some preliminary analysis that may lead to possible refinement of this experi ment;
IN TELLIG EN C E in D ESIGN CONTROL OF EXECUTION IN TELLIG EN C E in A NALYSIS
Model C onsid eratio n ( 1 )
" (2)
T a sk Ide ntification
Figure 4.2 : Extended model of Intelligent reasoning
(7)- evaluates this experiment;
(8a)- returns and executes another experiment that was already conceived as part of a glo bal strategy;
(8b)- reacts to evaluation of previous experiment and then consider further experimenta tion;
(9b)- exit from this cycle and proceeds;
(10) - analyses results; (11) - presents conclusions.
(iv) Decision-Making
In the context of the support framework envisaged in this chapter, where the machine pro vides the intelligence and the human executes the experiment, the support for the user in the decision-making process occurs at two levels : in the specification of experiments and in the analysis of the investigation. In other words, the output from the proposed advisory system must be a set of instructions on how to set up an experiment and advice about the conclusions that can be drawn after the experiment The user does not necessarily have to be informed about the higher level strategies which determine the design of the individual experiments.