Unsupervised Constraint Acquisition

Learning constraints through partial queries

“…However, a noticeable difference is that if there exists a constraint rejecting one positive example e + and also rejecting many (i.e., more than 100 in the standard setting) negative examples, Conacq.1 discards all networks containing this constraint from the version space whereas BayesAcq considers e + as an error. Prestwich also proposed SeqAcq, another passive learner robust to errors [36]. SeqAcq uses a statistical approach based on sequential analysis.…”

“…Research on constraint acquisition tackles this bottleneck. In most of the constraint acquisition systems (e.g., Conacq.1 [10,12,17], ModelSeeker [7], Arnold [29], SeqAcq [36]), the user provides examples of solutions (positive) and non-solutions (negative). Positive examples can come from historical data whereas negative examples can come from failed executions of attempted solutions.…”

Conversational Modeling for Constraint Satisfaction

“…The suggestion here is that as the acquisition process proceeds the constraint programmer might also learn how to solicit such knowledge directly from the domain expert, or how better to utilize volunteered information. SEQACQ (Prestwich 2020) is able to cope with errors in a training set, where instances are misclassified as solutions or non-solutions. The constraint programmer could use information acquired during the acquisition process to query the domain expert about errors or inconsistencies, e.g.…”

Active Disjunctive Constraint Acquisition

“…However, ModelSeeker can only find constraints from the catalog that hold on the specific structures it can recognize. Orthogonal approaches have also been proposed to perform error-resilient acquisition (Prestwich 2020;Prestwich 2021). Unlike previous approaches, trying to classify all examples, such methods consider that some examples are errors and eliminate them.…”