Coevolutionary Approach to Extracting and Predicting Linked Sets of Complex Decision Rules from Activity Diary Data

Abstract

A new approach for extracting and predicting decision rules for linked choices is developed and explored for use in computational process models of activity-scheduling and activity-profiling decisions. This coevolutionary approach, represented by a new agent, coevolutionary heuristic algorithm for deriving rules from N-dimensional travel activity links (Chantal), has been implemented in the ALBATROSS model. The algorithm induces sets of decision rules by incorporating the observed choice for all other facets as conditions in the decision table formalism. In the predictive stage, probabilities of activities are iteratively updated until a predefined convergence level is reached. The performance of the coevolutionary approach is compared with independent and sequential modeling approaches. The coevolutionary model turns out never to perform worse and often performs better than the best alternative approach under various conditions.