Seminar by Mateen Shaikh

Please Note: This seminar will be given VIRTUALLY.

Some Methods of Identifying Optimal Parametric Constraints

Model parsimony through parameter constraints are often found through search-and-score algorithms where many models are fit. This can be expensive and is often exacerbated by non-differentiable penalties/constraints, which disqualify many numerical algorithms. I'll give examples of two approaches that, while remain expensive, are at least simple for implementation. One student's approach of identifying model constraints via evolutionary algorithms. Another approach is anathema to statistical instinct by explicitly overparameterizing to obtain parsimony. This second approach continuizes discrete problems and employs sequences of infinitely differentiable functions for optimization. Sketch proofs of some minor lemmas of convergence as also provided.