This paper presents and evaluates sequential instance-based learning (SIBL), an approach to action seelction based upon data gleaned from prior problem solving experiences. SIBL learns to select actions based upon sequences of consecutive states. The algorithms rely primarily on sequential observations rather than a complete domain theory. We report the results of experiemnts on fixed-length and varying-length sequences. Four sequential similarity metrics are defined and tested: distance, convergence, consistency and recency. model averaging and model combination methods are also tested. In the domain of three no-trump bridge play, results readily outperform IB3 on expert card selection with minimal domain knowledge.
展开▼
