IBHYS: a new approach to learn users habits

Abstract

Learning interface agents search regularities in the user behavior and use them to predict user's actions. We propose a new inductive concept learning approach, called IBHYS, to learn such regularities. This approach limits the hypothesis search to a small portion of the hypothesis space by letting each training example build a local approximation of the global target function. It allows to simultaneously search several hypothesis spaces and to simultaneously handle hypotheses described in different languages. This approach is particularly suited for learning interface agents because it provides an incremental algorithm with low training time and decision time, which does not require the designer of the interface agent to describe in advance and quite carefully the repetitive patterns searched. We illustrate our approach with two autonomous software agents, the Apprentice and the Assistant, devoted to assist users of interactive programming environments and implemented in Objectworks Smalltalk-80. The Apprentice learns user's work habits using an IBHYS algorithm and the Assistant, based on what has been learnt, proposes to the programmer sequences of actions the user might want to redo. We show, with experimental results on real data, that IBHYS outperforms ID3 both in computing time and predictive accuracy.