Incremental Constructive Induction: An Instance-Based Approach

s feature construction process is not adequately constrained. Some application domains are described by adequate attributes, in which case feature construction could cause unnecessary search that decreases learning performance. 5 CONCLUSION This paper introduced and empirically evaluated IB3-CI, an instance-based learning algorithm that constructs features. It recorded higher accuracies and lower storage requirements than previous IBL algorithms , including MBRtalk (Stanll & Waltz, 1986), in two applications. IB3-CI is an incremental algorithm. This distinguishes it from several other recently introduced algorithms that support feature construction , which are non-incremental or require considerable pre-processing of instances to learn appropriate constructed features (e. incremental feature construction process allows it to perform well in the tic-tac-toe endgame experiments. IB3-CI is only an initial attempt to integrate constructive induction in IBL algorithms; its biases severely limit its applicability. However, its current level of performance suggests that additional modications are well worth investigating.