Open Forum Discussion:
This is a summary of the question and answer session between the audience and panelists at the 15th meeting. As always, the discussion was intriguing and too short. This summary cannot compare to being there in person.
Q: Addressed to GH - What design do you have in mind when you say only math models work?
- GH: Any design. If a design is rational it can be modeled mathematically.
- DS: Math is not necessarily a closed form solution, just like design.
Q: Addressed to GH - Can a descriptive model be used for design?
- GH: If you predict the future, the model needs to be predictive.
A discussion about TRIZ took place centering specifically on the validity and success of TRIZ.
- GH: With many methods there are successes but there are also failures.
- MW: Failure to find a counter example is not a proof.
- DS: Look at plurality voting. Mathematically this is not the best system yet we continue to use it. Looking at the past record of successes does not guarantee that tomorrow will bring the same success.
Q: What do you consider a proof?
- GH: A mathematically derived proof.
- RH: Argues GH's response by pointing out that "F = ma" is not mathematically proved.
Q: Can we really validate a model? Do we care about usefulness or do we care about validation in engineering? Can we find a quantitative assessment for validation? What about confirmation vs. validation?
- RH: When a model is proposed, someone should try to poke holes in it through experimentation. An experiment is useful when it tells you something you didn't know before. Experimentation can help improve a predictive model.
- DS: That's the scientific principle. We need to validate truly predictive models for putting things on the market. We need to create guidelines for validation and account for biases and errors. Validation is really playing devil's advocate.
- GH: Models aren't right or wrong but you can still make a tragic mistake using them. We need models that are consistent with what you want and know and this is not trivial. You can't get validation (for predictive models) from data, period.
Q: How can you develop guidelines for validating operational models? These are models that include decision making, and traffic laws, reality in general.
- GH: Predictive models cannot have feedback data.
- DS: To find out our weaknesses in validation we cannot just include future tests. We need to find weaknesses in advance - how can we test nuclear weapons? Finding weaknesses will help us find where we can make tests.
Q: How can we separate decision making from validation?
- DS: We need to find the best practices now - cannot take the easy way out.
Q: What evidence could I give to show that my method for generating designs is valid?
- GH: Derive it from beliefs - look at von Neumann's three axioms of existence proofs to validate.
- DS: No existence proofs exist because there are no real principles for validation.
Q: Has the scientific method been reduced to some guidelines to follow for engineering design?
- DS: Engineering is an art form, science is long term - engineers need to get something done quickly.
- MW: Need to be consistent but need to look backward in order to validate. We need to be consistent and consistency is hard to come by.
- GH: Design is not atomistic. We need to integrate our knowledge and make decisions that are consistent with nature and strive to reduce the inconsistency in our models.
Q: The people who use the model don't necessarily create the model? How do we coordinate the fact that people who use the model might use it differently than the people who created it.
- GH: It's your model if you use it.
Q: Should we maximize or minimize the information that a model provides?
- GH: Information is subjective, so quantity is irrelevant. Quality of the information is the important thing. How well does it support decisions? Suh's axioms are not axioms, they are a preference structure.
- DS: Need to understand where we can use crude information and where we need more refined information in our validation. This is some of the basis of my work. This is really a cost effectiveness issue and we need some general guidelines for information content.
- MW: A key issue here is that we need to define information.
Q: How does Suh compare to von Neumann?
- GH: von Neumann is trying to map options onto a line that will put them in a particular order, while Suh is using axioms to impart a preference structure on the designer.
 |