Position Paper: AID'00 Machine Learning in Design Workshop


Always Expect the Unexpected!


Introduction

The observation that design is a ``social'' activity has led researchers to study multi-agent design systems (MADS), where each ``agent'' is itself a knowledge-based system that works on its own portion of the design problem. An agent has autonomy, knowledge, goals, and the ability to coordinate and cooperate with the other agents. A multi-agent design system is built up of a collection of agents that each have design knowledge, often with some shared, centrally accessible place to record the design requirements and the developing design. The collection of agents may have an organization imposed on them at the time of the system's creation, or they may devise organizations dynamically as needed.

The need to reflect the fact that in real life a collection of cooperating designers don't always do the same thing every time -- i.e., that they learn -- led to the integration of Machine Learning (ML) techniques into MADS. Machine Learning is a subarea of AI that studies the knowledge and reasoning needed to make knowledge-based systems learn to perform better. One would expect learning to improve the efficiency, or effectiveness of the designers, both individually and as a whole: i.e., "better, cheaper, faster", one of design's mantras.

In [Grecu & Brown 1998] we discuss some Dimensions of Machine Learning in Design. There are many things that might be learned, and many things that can trigger learning.

Most previous approaches to learning in MADS take an existing ML technique and add it to an existing system so that it always learns the same sort of thing, in the same way, in the same place, at the same time. This is too inflexible.

It's clear that humans learn when there is a `need': often when they realize that some required information is missing. In order to handle the complexities of working in a group, a human designer often forms expectations about the possible reactions of other group members to his or her decisions. They may also learn approximations or default values for data that they need when making decisions. These too are expectations: expectations about what those values would be in that situation if they had in fact been available.

Expectations

In AI we are most familiar with expectations that are expressed using structured knowledge representations such as Frames [Minsky 1975] and Scripts [Schank & Abelson 1977]. These representations refer to known situations containing known things: for example, things playing roles, taking part in known relationships, or describing properties.

The slots in such representations were intended to provide strong guidance during `recognition' tasks. They could be used to recognize valid examples of a design, or a configuration, for example. By having strong expectations about what data might be appearing, the task can be made much more focused, and goal-directed: hence, more efficient.

While this sort of recognition-guidance use of expectations can be very powerful and, for example, forms the basis of description logic approaches to configuration [McGuinness & Wright 1998], it requires a lot of centralized prior knowledge. As one moves to less routine or to non-routine design and configuration problems [Brown 1996], a lot less will be known about the solution a priori. With non-routine problems the chances of getting the process or the result right the first time will decrease, and there is more need for learning [Brown 1998].

Multi-agent design systems add another dimension to this need to learn, as each agent doesn't know what the other agents know, nor do they necessarily know all the other agents. Expectations are needed in MADS as agents do not have `complete' knowledge about the designing process. Hence their view of the multi-agent system, and the designing knowledge it contains, is only partial. Hence, expectations, especially if they can be confirmed through successful use, allow an agent to act with more confidence of success.

Taking the LEAD

Grecu's work on flexible learning in multi-agent design systems allows the agent system to decide what to learn and when to learn it [Grecu 2000] [Grecu & Brown 2000]. This requires the detection of a need for learning, the investigation of what information might be used as a basis for learning, the gathering of suitable data of that type in order to form a Training Set, the process of learning from that data, and the validation of what was learned.

In Grecu's LEAD system, agents either make a design decision, or criticize another agent's decision. Critics can reject a prior decision or offer advice about a better value. The design agents can learn expectations about suitable values -- if values that they need are missing -- or expectations about the consequences of the agent's decision with respect to future rejection or criticism.

Expectations are learned by an agent when it detects that it often has missing data or when its decisions are often subsequently treated negatively by critic agents. At that point the agent does a `causal analysis' in order to try to determine what features of the situation might have contributed to the problem. It uses knowledge of the domain and of its own use of data in order to come up with a set of candidate features.

The system then experiments with similar design requirements in order to produce a set of data (i.e., real examples of those features, such as X = 6 ). This data, and the candidate features, are fed to a machine learning technique known as a Wrapper, that uses inductive learning to reduce the set of features to the set that is most useful for forming an expectation. i.e., the most predictive set.

An expectation is then formed from those features -- for example: 

         IF  The frame material is {aluminum, aluminum alloy}
            & The seat thickness is between 1.5 in and 2 in
         THEN  The chair price will be in the range (90,110)
The system validates the expectations using other test cases, and then monitors its use while designing.

Its Significance

The significance of this flexible learning approach is that each agent can respond to the actions of other agents that affect it, even if they are added to the system, and can adjust over time as new interactions are discovered as design requirements are changed. The system even reacts to the learning of other agents: new expectations may cause different decisions, thus affecting other agents, and causing them to produce new expectations.

The key aspect of this learning technique is the knowledge-based, context-dependent causal attribution process, where an agent that is trying to learn will reason about its current context in order to extract potential causes of the problem it is facing. This is an original contribution of this work. The use of the wrapper to prune the potential causes to those most predictive is another original contribution.

This work is significant as any method that serves to improve the design process has significant potential for commercial use, and, given the importance of designing to our economy, significant economic impact. There is also significance in the fact that this is a new learning method for multi-agent systems. It's also the first to use expectations in this way. It is also significant because its flexible approach is tightly coupled with the fact that it applies to design systems in particular, and because it achieves a very tight integration with designing, in contrast to many methods that are dependent only on the multi-agent context. Consequently, this research has both high practical and theoretical significance.

References

D. C. Brown, ``Routineness Revisited'', Mechanical Design: Theory and Methodology, (Eds.) M. Waldron & K. Waldron, Springer-Verlag, 1996, pp. 195-208.

D. C. Brown, ``Improving Design with Agents, or, Improving Agents by Design'', Workshop on Intelligent Agents and Their Potential for Future Design and Synthesis Environment, http://www.cs.wpi.edu/~dcb/Papers/UVA-NASA-paper.pdf , NASA Langley Research Center, Hampton, VA, Sept. 1998.

D. L. Grecu, Flexible Learning in Multi-Agent Systems, Ph.D. Dissertation, CS Dept., WPI, Worcester, MA 01609, May 2000.

D. L. Grecu & D. C. Brown, Expectation Formation in Multi-Agent Design Systems, AID'00: the 6th Int. Conf. on AI in Design, June 2000.

D. L. Grecu & D. C. Brown, Dimensions of Machine Learning in Design, Artificial Intelligence for Engineering Design, Analysis, and Manufacturing Journal, special issue on Machine Learning in Design, (Eds.) A. H. B. Duffy, D. C. Brown & A. K. Goel, Cambridge U.P., 1998, pp. 117-121. {A summary is available at http://www.cs.wpi.edu/~dcb/AID/AID96/taxonomy.html }.

D. L. McGuinness & J. R. Wright, ``Conceptual Modeling for Configuration: A Description Logic-based Approach'', Artificial Intelligence for Engineering Design, Analysis, and Manufacturing Journal, special issue on Configuration, Vol. 12 No. 4, Sept. 1998, pp. 333-344. {Available at http://www.journals.cup.org/owa_dba/owa/approval?sjid=AIE&svid=12&siid=4&spii=S089006049812406X  }.

M. Minsky, A Framework for Representing Knowledge, Psychology of Computer Vision, P. Winston (Ed.), McGraw-Hill 1975,

R.C. Schank & R.P. Abelson, Scripts, Plans, Goals and Understanding: an Inquiry into Human Knowledge Structures, Lawrence Erlbaum Associates, 1977.


Version: Mon May 22 17:51:18 EDT 2000