The graph of arguments and attacks is a product of the knowledge base, and so any changes to it have to be expressed as changes to the knowledge base. This is why we imposed a number of constraints on the ASPIC+ systems used in this thesis (see Section
2.3.2 on page 32). Our contribution to the field is therefore to establish the parameters that allow enforcement in ASPIC+.
3.5.4.2
Dynamics of Argumentation Systems
Our development of enforce and related operations was driven by the requirement to change the status of a set of arguments from credulously accptable to sceptically ac- ceptable. The change of status of an argument after updating an argumentation system has been researched under the term argument dynamics [84], or dynamic argumenta- tion frameworks [85]. Similar to the work on belief revision discussed above (a lot of builds upon argument dynamics), these approaches focus abstract argumentation or DeLP [86, 87]. While our work is aimed at (a subset of) the ASPIC+ system, it would
be interesting to compare it with prior work on DeLP and ultimately position it in the framework of belief revision, similar to the approach taken in [81].
3.6
Conclusion
We will now discuss how ADF meets the five requirements that were set in section 3.2.4, and then review how the developments of this chapter prepare us to analyse decision processes, sequences of decisions, in the next chapter.
3.6.1
Discussion
In Section 3.2.4 we laid out five shortcomings of the traditional, non-argumentation based approach to decision making in light of knowledge reusability. Our decision framework ADF addresses these five points:
1, Opaque resoning process Preferred decisions in ADF are backed by arguments. Those arguments are part of grounded extensions of an argument graph generated from a knowledge base and are therefore based on a formal model of the domain. This model is part of the ADF, which means outside of that knowledge base, no additional information is needed to reproduce the reasoning process.
3.6. Conclusion 106 2, Local optimum With the notion of recommended decisions in ADF, a decision maker does not have to enumerate the possible decisions manually. Rather, they are the decisions are defined by the knowledge base. The problem of identifying possible decisions has thus been replaced by the problem of modeling the domain.
3, Proprietary documentation formats Since ADF relies on a formal language with an inference mechanism, it may easily be transformed to any logic-based model of knowledge, for example an ontology. Ontologies, which are based on description log- ics, are a widely used method of recording domain knowledge. Conversely, domain knowledge from an ontology can be transformed into ADF rules. Williams and Hunter [25] describe how knowledge from an ontology can be used in the argumentation pro- cess.
4, Manual analysis With the domain knowledge represented as rules used in argumen- tation, automated decision analysis is possible as discussed in Sections 3.3.3 and 3.3.4. 5, Costly retrieval of documentation For ADF systems, the problem of re-evaluating previous decisions depends mainly on the data structure chosen for the knowledge base, because it is sufficient to retrieve the relevant rules and assumptions from previous decisions. It is not necessary to perform the entire analysis again.
In Section 3.2.3, we presented decision reusability and claims management as two use cases for documentation in the aerospace industry. Both require a formal, structured documentation of decisions. If an argumentation based system is used for decision making, then the documentation could be generated by the same system with no additional cost, since the arguments used by the system to evaluate decisions are, at the same time, formal justifications of decisions. Because they are structured (as determined by the contents of the knowledge base), they can easily be converted into a structured format for documentation, for example an ontology.
In this chapter we developed an argumentation-based model of decision making that combines decision making with uncertainty with multi-criteria decision making. We now have the building blocks needed to model sequences of decisions: We can evaluate options according to how well they meet our criteria under scenarios, and we can adjust our knowledge base after making a decision.
3.6. Conclusion 107
3.6.2
Future Work
The interplay between the logic based on which arguments are formed, and the ac- ceptability of arguments – as determined by argument graph semantics – looks to be a promising area for future work in argumentation in general [88], and this is also the case for our application of the theory in the domain of decision making.
We see opportunities for future work mainly in the area of decision making with uncertainty (Section 3.3.4). First, we want to extend our notion of possible worlds (Definition 32) with a degree of probability - whether qualitative as in the work of Dubois [75], or quantitative, using classical probability theory or one of the weaker alternatives proposed in the AI literature. Since possible worlds in our understanding are the same as preferred extensions and thus sets of arguments, assigning probabilities to them may create an interesting dynamic with the underlying knowledge base. This work can build on recent proposals for probabilistic argumentation (e.g. [60, 61]).
A second and related concern is the relationship between probabilities of possible worlds, and the utilities of goals. In the purely symbolic setting of [54], this relationship was the source of some problems related to the consistency of the system’s output. It would be interesting to investigate if the fact that our decision problems are generated from a knowledge base allows us to solve some of those problems, based on the fact that arguments (and extensions) are internally consistent.
In the area of multi-criteria decision making (Section 3.3.3), we want to improve the definition of satD (Definition 29), the function determining how many criteria are
met. In this thesis, sat simply returns a set of criteria that are achieved by a given option, and does not distinguish the criteria any more. Therefore, the notion of domi- nated options (Definition 30) can only be applied if an option O satisfies all of another option O’s criteria, which seems to be relatively rare in practice. We want to consider additional means of summarising the outcome of an option, similar to the way an ag- gregation function agg can be defined in the traditional approach to MCDM (Definition 23).
In the area of enforcement and deactivation of rules, we are planning to compare our proposal with the work on enforcement in abstract argumentation [65, 66, 67, 68], and to position them within the wider field of belief revision.
Chapter 4
Argument-based Decision Process
4.1
Introduction
In Chapter 3 we developed an argumentation-based model for decision making, tak- ing into account the need to balance several criteria as well as uncertainty about the outcome of options. In our framework we also have the ability to adjust a knowledge base once a decision has been made, so that knowledge which used to be only one of several options becomes part of the “safe” core of the knowledge base (the grounded extension).
Let us now take a step back and take another look at our goals set out in the intro- duction. We want to describe not just individual design decisions, but entire processes consisting of many decisions.
A lot of uncertainty exists in design processes [6], not only on the level of indi- vidual decisions (which we considered in Chapter 3), but also on the process level – for example, it is not clear what requirements will exist at the end of the process, be- cause they change frequently. For example, the design of a new car should combine new technological advancements (efficient engines, computer-assisted driving and so on) with the aesthetic language that has already been established as part of the car’s brand. It also has to meet a target price and profit margin. It is common for decisions to be reconsidered at later stages, when more knowledge has become available. In the car example, new knowledge could take the form of changed commodity prices or tax breaks in some countries.
In this chapter we will extend our decision model by looking at decisions in con- text, taking decision frames (see Chapter 3) as the building blocks of decision pro-
4.2. Use Case 109