The First Workshop on Advances in Evolutionary Computation (WAEC 2014) is a one-day meeting that will present current progresses on evolutionary computation and related areas. We intend to provide a framework for fruitful and informal discussion on current state of the art topics in EC.
WAEC is organized under the scope of the German DFG Collaboration project “Addressing Current Challenges in Evolutionary Multi-Objective Optimization: Many Objectives, Indicator-based Selection and Convergence” involving researchers from the University of Münster, TU Dortmund and PUC-Rio and the CNPq BJT project 407851/2012-7.
Empirical surrogate models which predict the workpiece properties and the process characteristics resulting from a specific setup allow the potential of the available resources to be explored. Based on a set of initial experiments, the properties produced by parameter settings not yet conducted can be predicted. In particular, the use of kriging models from the Design and Analysis of Computer Experiments (DACE) has shown many successful applications over the last 20 years.
In this talk, a framework for planning manufacturing processes is presented. This framework is based on kriging models from DACE and two process chain optimization approaches. As a part of this framework, a procedure for the model-based analysis and optimization of manufacturing processes is proposed. Enhancements with respect to the presence of noise in the empirical data are discussed and the resulting improvements are evaluated using simulation studies. Moreover, novel sequential design criteria for approximating the Pareto frontier, i. e., the set of optimal compromises with respect to the workpiece properties or process characteristics, are introduced and assessed in a theoretical and empirical manner. They allow the possibility of a local refinement provided by the kriging models to be exploited. The Pareto frontier is of particular importance, as it represents a compact overview of the potential of the corresponding manufacturing process and comprises the parameter combinations of interest for the manufacturing planner.
The validity of the planning framework and the model-based procedure is documented based on a thermomechanically coupled process chain for manufacturing self-reinforced thermoplastic single-polymer composites. The formalization of this process chain by means of a collection of empirical models is motivated and presented. A particular focus is put on assessing different possibilities to predict the spatial distributions of workpiece properties measured by performing impact tests on specimens locally prepared from the resulting component. It is shown that this process chain can be accurately modeled, analyzed, and optimized using the proposed framework.
How a Decision Maker (DM) should behave when unaware of the relative importances of the stated decision criteria and of the future consequences of his/her actions?
The traditional premise in sequential Multi-Criteria Decision-Making (MCDM) under uncertainty is that all relevant information for characterizing the problem is readly available. This is however an unrealistic statement in practical scenarios, wherein: (1) the system must resort to noisy historical observations for model building and parameter estimation; (2) the DM has little knowledge about the underlying trade-offs of the problem; and (3) the DM has little subsidies for deciding whether near or long-term performance is preferable.
In this talk, we argue that eliciting complete preferences and eagerness to near-term optimized performance under little knowledge can be unattainable at best and too much speculative and deceptive at worst. We will thus present research on Anticipatory Stochastic Multi-Objective Optimization (AS-MOO) and sequential MCDM systems capable of simultaneously handling challenges (1)-(3) requiring minimal involvement from the DM. The anticipatory capabilities of multi-objective evolutionary algorithms are thus augmented with Bayesian models so as to approximately solve the formulated AS-MOO problems.
Partial and often single-objective rules of thumb cannot easily be combined to multi-objective rule sets. Further, standard hybridization approaches usually increase complexity of methods and often necessitate a re-design of the algorithms which cannot be done by a standard user leaving him/her with unmodified standard approaches in the end. In this talk, we motivate the integration of available knowledge or expertise as a major challenge in bridging the gap between applicable single-objective scheduling rules and their application in multi-objective scheduling. We highlight this gap and briefly discuss the issues of standard approaches from evolutionary computation. As a first solution to these issues, we propose an agent-based approach inspired by evolutionary computation that may lead to a flexible framework for integrating single-objective expertise in a generic multi-objective solution strategy.
Predio Cardeal Leme, 4to andar
Sala Multimeios do DEE (entrada por sala 401L)
Rua Marquês de São Vicente, 225, Gávea
Rio de Janeiro, RJ – Brasil – 22451-900