Creativity by Case-Based Reasoning (CBR): SWALE Project Home Page

The URL for this page is http://www.cs.indiana.edu/~leake/projects/swale.

Project Description

The task in which the SWALE project studies creativity is creative explanation of anomalous events. SWALE is a story understanding program that detects anomalous events and uses CBR to explain the anomalies. Its explanation process is based on the retrieval and application of cases storing prior explanations, called explanation patterns from its memory. This case-based reasoning process can take place at any point along the spectrum of creativity, from the completely non-creative application of a perfectly appropriate XP, to the very novel use of an inappropriate XP that must be totally revised to be usable.

The following illustrations of SWALE's reasoning are excerpted and adapted from Schank, Roger C., and Leake, David B. Creativity and Learning in a Case-Based Explainer. Artificial Intelligence 40(1-3):353-385, 1989, and reprinted in Carbonell, J., ed, Machine Learning: Paradigms and Methods, MIT Press, Cambridge, MA, 1990. A simplified version of the SWALE code, developed in scheme for pedagogical purposes, can be downloaded and run to illustrate this process.

The Range of Explanations Generated by SWALE

In 1984, Swale was the best 3-year-old racehorse, and he was winning all the most important races. A few days after a major victory, he returned from a light morning gallop and collapsed outside his stable. The shocked racing community tried to figure out why. Many hypotheses appeared, but the actual cause was never determined.

The SWALE system detects the anomaly in Swale's premature death (see (Leake, 1992) for a discussion of the system's anomaly detection process). It then builds explanations of the death by retrieving and "tweaking" remindings of XPs for other episodes of death. It picks the best of these candidate explanations, and adds it to its XP library for future use. Some are reasonable explanations; others are quite fanciful or can be ruled out on the basis of other knowledge. However, they show that a memory-based explanation system, even if it has a limited range of XPs and of retrieval and tweaking strategies, can come up with a variety of interesting explanations.

• Reminding Thinking of other deaths of those in peak physical condition causes the system to be reminded of the death of the runner Jim Fixx, who died when his running over-taxed a hereditary heart defect.

  Explanation Swale might have had a heart defect that caused his racing to prompt a heart-attack.

• Reminding Thinking about other deaths of young stars, the system is reminded of Janis Joplin's death from a drug overdose.

  Explanation 1 The pressure of being a superstar was too much for Swale, and he turned to drugs to escape. He died of an overdose.

  Explanation 2 Swale might have been given performance-enhancing drugs by a trainer, and died of an accidental overdose.

• Reminding Thinking of folkloric causes of death causes the system to recall the old wives' tale too much sex will kill you.

  Explanation 1 Although racehorses are prohibited from sex during their racing careers, Swale might have died of a heart-attack from the excitement of just thinking about life on the stud farm.

  Explanation 2 Swale might have committed suicide because he became depressed when thinking about sex.

  Explanation 3 Swale might have died in an accident when he was distracted by thinking about sex.

Requirements for Creative Case-Based Explanation

The above examples show that interesting explanations arise when we try to use an XP that doesn't quite apply. In order to obtain creative explanations, an explainer might try to intentionally misapply XPs. Interesting ideas can arise from using old explanations to deal with situations where those explanations were never intended to be used.

While using an XP that doesn't apply gives a fresh perspective on a situation, the idea of building a system to intentionally misapply XPs raises many issues: Which XPs do you retrieve? Which tweaks should be applied? How long should the tweaking process be continued? As our research progresses, we hope to be able to answer these questions. However, we can suggest some of the things that are needed to build creative case-based explainers:

• We need heuristics for the intentional reminding of explanation patterns XP retrieval is the process of formulating questions to memory: we characterize an anomalous situation in terms of a set of indices, and ask what XPs in memory explain similar situations. When no answer is available, we must reformulate the question into one that we can answer. When no solution is directly available, people often fall back on asking standard questions that give background information. Answers to explanation questions like what physical causes underlie this event?, what special circumstances made the event happen now?, what motivates the actor of this surprising action?, how did the victim enable this bad event?, or what groups might the actor be trying to serve?, may suggest relevant factors that can be used as indices for XP retrieval. Though the XPs accessed in this way might not be directly applicable, it may be possible to adapt them. A creative system needs a set of explanation questions for gathering information, rules for selecting which questions to apply in a given situation, and rules for transforming them to fit.
• We need tweaking strategies that can do significant revisions Tweaking must also be able to make significant revisions. Rather than requiring tweaks to always maintain causal structure, we should allow them to make broad changes. Their revisions will not always be successful, but failures produce new possibilities for still more revision.
• We need heuristics for knowing when to keep alive seemingly useless hypotheses In addition to chosing between explanations generated by the system, the evaluation process also has a more direct part in the creative process. We cannot tweak a candidate explanation indefinitely; the evaluator must decide whether a hypothesis is worth pursuing. This estimation will always be imperfect, but the better it is, the more resources the system will be able to devote to fruitful revision of XPs. One heuristic would be to continue tweaking an explanation as long as each tweak generates a better explanation. But the decision whether to continue tweaking should also depend on the availability of competing candidate XPs, and on an estimate of how important the final explanation is to goals of the system (since that affects how many resources should be expended explaining).
• We need a system with a rich memory of explanations Finally, a creative case-based explainer must have access to a wide range of explanation patterns. There are two ways that people or machines might learn new XPs: by being taught them directly (as children are given explanation patterns by parents, teachers, or friends), or by learning new ones through creative misapplication. One step towards making a computer creative would be to collect an extensive list of XPs that it could use as the starting point for adaptation. Many interesting explanations might be constructed starting with a collection of culturally-shared XPs, such as proverbs.

Code On-Line

Sample References

• Schank, Roger C., and Leake, David B. Creativity and Learning in a Case-Based Explainer. Artificial Intelligence 40(1-3):353-385, 1989, and reprinted in Carbonell, J., ed, Machine Learning: Paradigms and Methods, MIT Press, Cambridge, MA, 1990.
• Schank, Roger C. Explanation Patterns: Understanding Mechanically and Creatively. Erlbaum, 1986.
• Leake, David B. Focusing Construction and Selection of Abductive Hypotheses, Proceedings of the Eleventh International Joint Conference on Artificial Intelligence, 1993, pp. 24-29. This paper describes the relationship of case-based explanation to other explanation approaches. View abstract only.

Project Team

Contacts