User:Blankjul/sandbox/Surrogate-Assisted Optimization

Surrogate-Assisted Optimization describes an optimization method making use of a so-called surrogate model to improve convergence behavior. Using the approximation of surrogates is a well-known technique for optimizing time-consuming and computationally expensive objective and constraint functions. The algorithmic overhead introduced by the usage of a surrogate assisting the search is (often) negligible compared to the overall dominating time for evaluating a solution.

Literature

Goal: Metadata, such as title or keywords of a publication, often focus on the problem to be solved and overall. In surrogate-assisted optimization, the goal is to converge rather quickly to an optimum by considering the global trend of a function as much as possible. Therefore, relevant publications are being found under the keyword efficient global optimization. The majority of methods proposed in the literature make use of auxiliary models and, thus, are, in fact, surrogate-assisted algorithms. Moreover, the overall goal is also related to the purpose of the so-called anytime optimization. As indicated by the name, the focus lies on methods that can be interrupted anytime, having achieved the best result possible. In an optimization context, this is commonly realized by a comparison of the algorithm's convergence curves during an optimization run to consider not only the final but also the result in each iteration. Depending on the concrete experiment, this usually favors quickly converging algorithms and ensures suitable results in the long run.

Method: Most of the time, the title and abstract of a publication reveal the techniques used to achieve a specific goal. Thus, the usage of surrogate used to accelerate the convergence is often explicitly mentioned since it is an inherent part of the algorithm. Nevertheless, researchers refer to a model most commonly as a surrogate; the terms metamodel, response surface model, approximation model, or simulation model have been used in the past. Although some terms might have (slightly) different meanings in general, one might find relevant publications by refining the search considering a different terminology.

Problem: Despite emphasizing the goal or method, the problem statement itself can be of more importance. Problem-related literature requires a search regarding different types of problems that are either computationally expensive or supposed to be optimized with a minimal budget of function evaluations. Typically, simulation-based problems are of an expensiveness nature and are thus worth to consider. For simulations, the evaluation time for a single solution might become a couple of hours or days. Moreover, one can think of optimization problems where the objective or constraint function needs to process a large amount of data. Often, the data have been collected empirically, and stochastic optimization is supposed to be applied. A literature review reveals that many existing studies use distributed systems to speed up the function evaluations, and only rarely surrogates have been used.