Searching for Novel Classifiers

Abstract

Abstract. Natural evolution is an open-ended search process without an a priori fitness function that needs to be optimized. On the other hand, evolutionary algorithms (EAs) rely on a clear and quantitative objective. The Novelty Search algorithm (NS) substitutes fitness-based selection with a novelty criteria; i.e., individuals are chosen based on their uniqueness. To do so, individuals are described by the behaviors they exhibit, instead of their phenotype or genetic content. NS has mostly been used in evolutionary robotics, where the concept of behavioral space can be clearly defined. Instead, this work applies NS to a more general problem domain, classification. To this end, two behavioral descriptors are proposed, each describing a classifier's performance from two different perspectives. Experimental results show that NS-based search can be used to derive effective classifiers. In particular, NS is best suited to solve difficult problems, where exploration needs to be encouraged and maintained.

Preliminary Study of Bloat in Genetic Programming with Behavior-Based Search

“…However, there is a trend, NS performs well on the more difficult problems, and worse on the easier ones. These results are similar to those reported in [24,23], with similar conclusions. Basically, the explorative search performed by NS is fully exploited when random initial solution perform badly, under these conditions the search for novelty can lead towards better solutions.…”

“…First, Table 2 presents the average classification error on the test data. The results are consistent with those reported in [24], NS-SRS performs well, with small performance differences compared to SRS. However, there is a trend, NS performs well on the more difficult problems, and worse on the easier ones.…”

“…Since its proposal in [15], and later works [16,17], most applications of NS have focused on robotics, such as mobile robot navigation [15][16][17], morphology design [18] and gait control [17]. Only until recently has NS been used in general pattern recognition problems, particularly supervised classification [24] and unsupervised clustering [23]. This paper will use the work in [24] as the case study to analyze bloat in a NS-based GP search.…”

“…Only until recently has NS been used in general pattern recognition problems, particularly supervised classification [24] and unsupervised clustering [23]. This paper will use the work in [24] as the case study to analyze bloat in a NS-based GP search.…”

“…This similarity measure has been used to measure behavioral diversity in ER [22]. It is noteworthy to point out that [24] reports good performance on several synthetic classification problems. In particular, the authors report a clear trend, the performance of NS-based GP improves, relative to a control method, as problem difficulty increases; i.e., NS performs comparatively better on hard problems than on easy ones.…”