A knowledge-based equation discovery system for engineering domains

KEDS (knowledge-based equation discovery system), which integrates machine-learning and statistical techniques to learn a range of comprehensible models in nonhomogeneous domains, is described. The intertwining of partitioning and discovery enables it to learn relationships from data and extract their underlying structure, and probabilistic clustering enhances runtime performance and improves accuracy. KEDS uses a divide-and-conquer strategy, breaking the engineering problem space into smaller regions so as to meet two goals: partitioning should make it easier to discover the models in each region, and he resulting model for each region should meet the comprehensibility requirements imposed by the engineer. KEDS is also a recursive fit-and-split system, which finds partial hypotheses (candidate equations), and then partitions the problem space into regions.<>