Conversion in Progress

Chapter 7
Potential Research Directions

As chapter six attests, some limited research efforts and a few "real world" applications have sought to synergize DS and AI techniques. Considerably more however can be done. This chapter looks in detail at what can be done to promote further synergism, and also looks at some additional, yet more practical ways that AI can be applied in DM.Promoting Further Synergism Five ways in which the AI, DS, and combined disciplines can be better served are outlined below.An understanding of commonality/extensionality should be promoted. As it now appears, considerable duplication of research efforts exist between the DS and AI research communities. Additionally many of the research techniques being developed duplicate processes or produce results already thoroughly studied or outlined in other disciplines. More coordinated planning by the government agencies promoting this research as well as cooperative efforts by the major professional societies should be encouraged. This would include such efforts as joint meetings, or symposia devoted to technique comparison.Better modeling environments should be created and standardized. This need seems to be one expressed by a number of technology forecasters. For example in the July-August 1988 issue of Operations Research, the Committee On the Next Decade in Operations Research (CONDOR) proposed 16 areas targeted for research in the 90's. Eight of the sixteen directly addressed improved modeling environments. These suggestions included "Incorporation of intelligent user interaction facilities into a model", "Extension of the concept of a model to involve inference", and "Embedding learning capabilities in models and in systems that manage models." Fordyce et al in INTERFACES, spelled out needs in a similar manner. Their focus was upon improving our ability to; 1) manufacture situations requiring models based on heuristics, 2) deliver guidelines delivered from models, 3) imbed methods for choosing and formulating an appropriate model, and 4) interpret model results. A reasonable, practical way better environments could be aided would be through the search for a Goal, Criteria & Alternatives. Various reasoning methods could be used to first help establish a model. Some current software packages (e.g. Decision Maker[]) use a process approach toward eliciting & structuring the problem. However there is no formal inherent guidance for unfolding specific problem types, such as those covered in Chapter 2. Nor is there any method for guiding the user into discovery of similar "real world" models, i.e. analogical reasoning. At the same time search algorithms and heuristics could be used for the discovery, access and organization of both internal and external data and/or knowledge sources.A greater emphasis on the qualitative aspects of problem solving and decision making should be encouraged. A firm definition for qualitative remains elusive. Nonetheless we can establish a baseline which includes the myriad of attributes or qualities surrounding an object or event, and the non-numerical relationships among those attributes. Karl Weick at the University of Texas has formulated 66 primary "Thinking Strategies" which focus upon approaches toward problem solving. These strategies could be thought of as a list of possible relationships among objects or events. A portion of two of these strategies are presented below: 

Strategies for:                                                            

    Involvement Information                                                

     Committ                              Check                            

     Defer                                Diagram                          

     Hold Back                            Display                          

     Leap In                              Organize

In the same manner, the thrust of the CYC project discussed in Chapter 4 is to include "reasonable" approaches toward understanding the relationships among objects, events and their attributes. Other efforts generally classified in the AI discipline are now enabling progress in this area. These include object oriented systems and qualitative physics. A practical implementation in the qualitative area is the use of knowledge based subject matter "front ends" and deamons in systems. Extending the reasoning above, once recognizing entry into a specific domain or within a certain topical area, KBSs could be utilized to help guide the DM through the peculiarities of that area. E.G. for a budget decision, guidance could be made available concerning organizational policy & regulations. Or detailed information on cost areas such as CBA, types of costs, and cost measurement would be readily apparent. Yet another area ripe for practical implmentation is Conflict Recognition (not resolution) in subject matter content. A capability is needed to identify areas where events are being considered in a mutually exclusive sense when they are in fact codependent, or perhaps inconsistent in a rational sense. Examples:ÚÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄ¿³With regard to size: ³ ³ A is greater than B ³³ B is greater than C ³³ C is greater than A < cannot be true ³³ if 1 & 2 are true ³ÀÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÙ ÚÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄ¿ ³Proposal: ³ ³ ³³Lets reduce the deficit by decreasing the tax rate ³ ³to 10 %, while increasing spending by 25%. ³ÀÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÄÙ A growing area created for handling these type of problems in AI is called Truth Maintenance Systems (TMS). Goals of TMSs include recognizing problems such as these encompassing contradiction, redundancy, and circular reasoning.Consistency measures already present in some software such as Expert Choice should be propagated. Often the best that can be done for dealing with more qualitative attributes of problems is to simply recognize the complexity of many objects and situations and to plan for the unexpected within sophisticated modeling environments. Fortunately this can be aided considerably by providing more flexible and powerful capabilities to these environments.Practical methods for interfacing should be implemented. Immediate improvement in creating effective, combined DS/AI environments could be seen by creating better transitional forms for data, objects, and their attributes, whether quantiative or qualitative. While de facto standards of some types now exist on personal computers, they are limited. These include standards for worksheets such as Lotus 1-2-3's .WK1 files, for databases through dBase III+'s .DBF files, and a common graphics format via .TIF files. Other efforts which should be encouraged include the move towards standards in dynamic data exchange (DDE), as well as efforts within IEEE in its AI Standards Committee. M. Geneserath of Stanford University has created a standard for interfacing between knowledge bases called GIK. Hopefully this standard will also bear fruit.The idea of "Intelligent Matching" of problems and decisions to techniques for resolution should be expanded. As discussed in Chapter 3, problem solvers and decision makers now tend to frame problems and to solve them based upon their experience, whether valid or not. Additionally as discussed in Chapter 6, problem solvers also become involved in methodoliatry, solving all problems using one approach. Intelligent matching, the use of the computer to help select a method (or methods) and to guide the user through its use would go a long way toward rectifying these problems with current decision making. Intelligent matching uses attributes of a problem to suggest methods for resolution, along with details on how these can be implemented. User's would be "trained" in the method as they proceed through its implementation. Andriole[] has suggested a task/user/organization 3D matrix for guiding selection of a technique. Applegate [] has suggested defining problems along this same general line using task/data/situation/implementation characteristics. She has implemented a system, labeled "Method Master" in an object oriented environment on the Apple MacIntosh. Her approach extends a decision decomposition protocol proposed by Zachary []. Banerjee and Basu have proposed an AI frame based structure for adding intelligence to the matching process. And finally Juell, Nygard, and Nagesh have suggested a similar approach using neural network technology as the basis for selection of a method.

Paper Summary and Conclusions

The early part of this paper focused upon definitions for problems, decision making, and modeling. This was expanded using a cognitive basis with the purpose of pointing out the limitations and biases of human decision making. Technologies in the decision sciences and artificial intelligence areas were then introduced with the general purpose of demonstrating which weaknesses they supported and which biases they mitigated. The power gained by combining many of these techniques was presented. At the same time overlap and duplication in the thrust of the AI and DS disciplines suggest a problem in research directions. Finally five areas where improvement in AI and DS models can be realized was detailed. As a bottom line synergism between DS and AI can be improved. Professional level policy adjustments and the incorporation of AI and DS into DM through smoother transitional means are two paths to ensure this progress.