NAFIPS 2005 - 2005 Annual Meeting of the North American Fuzzy Information Processing Society

Fuzzy Evolutionary Systems:First Steps Towards Supporting Model Lifecycle

Piero P. Bonissone

General Electric Global ResearchOne Research Circle, K1-5C32ANiskayuna, NY 120309, USA

Email: bonissone@crd.ge.comVoice: +1 518 387-5155; Fax: +1 518 387-6845

Abstract

The typical lifecycle of a knowledge-based model starts from its development, testing, optimization, anddeployment, and continues with its maintenance phase. The latter consists of monitoring the model'sperformance, editing its knowledge base to prevent obsolescence, and updating the model when required.Quite often, however, models are handcrafted, i.e., a large amount of manual intervention is used in theearlier phase of their lifecycle. This leaves the maintenance phase as an "after-thought", often requiring asimilar level of manual efforts. We propose a process, based on evolutionary algorithms, that follows themodel throughout its entire lifecycle. For deployment, it generates a collection of competing models,evaluates their performance in light of the currently available data, refines the best models usingevolutionary search, and after a finite number of iterations, generates the best-found model. This process isrepeated periodically to automatically produce new updated versions of the model.

We chose an asset selection problem to illustrate this methodology. Given a fleet of industrial vehicles(diesel electric locomotives), we want to select the best subset (of fixed or variable size) for mission-criticalutilization. To this end, we predict the remaining life for each unit in the fleet. We then sort the fleet usingthis prediction and select the highest ranked units. The model chosen to perform this prediction/selectiontask is a fuzzy instance-based model. Unlike functional approximators that require an off-line supervisedlearning stage to create a hyper-surface in the cross-space state-output, instance-based models (IBM's) onlyrequire accessing the data in a local neighbourhood of the new point defined by the query. IBM's rely on acollection of previously experienced data stored in their raw representation. Unlike Case-Based Reasoning(CBR), they do not need to be refined, abstracted and organized as cases. Like CBR, IBM's represent ananalogical approach to reasoning since they rely on previous instances of similar problems and use them tocreate an ensemble of local models. Hence the definition of similarity plays a critical role in the theirperformance. Typically, similarity will be a dynamic concept and will change over the use of the IBM's.Therefore, it is important to apply learning methodologies to define and adapt it. Furthermore, the conceptof similarity is not crisply defined, creating the need to allow for some degree of vagueness in itsevaluation. Hence, we propose the use of Fuzzy IBM's (F-IBM's). We address the issue of similarity byevolving the design of a similarity function in conjunction with the design of the attribute space in whichthe similarity is to be evaluated. Specifically, we use four steps: (1) Retrieval of similar instances from thedatabase (DB); (2) Evaluation of similarity measures between the probe and the retrieved instances; (3)Creation of local models using the most similar instances (weighted by their similarity measures); (4)Aggregation of outputs of local models to probe. Within the example of asset selection, we show theaccuracy of the evolved F-IBM's, their robustness to information loss, and the benefit of their automatedupdating process to avoid performance loss. Finally, we advocate the use of evolutionary searchintertwined with local search to fuirther improve model life cycle.

Keywords: Fuzzy Similarity, Evolutionary Learning, Instance-based models, asset selection.

0-7803-9187-X/05/$20.00 ©2005 IEEE. 579