This talk summarizes our 15+ years of work on the use of Machine Learning for Search & Optimization. I review the four main approaches that we invented during this time. Since learning during search takes effort, it should not surprise that we designed three of these approaches for a particular target range of total function evaluations: from few tens of dozens, to thousands, to many hundreds of thousands of function evaluations. The last hybrid I review regards a surrogate-based approach for optimization under stochastic uncertainty. The wonder of this research area is that each of these four methods defines the state of the art in its respective area, giving significant empirical evidence that learning to optimize can be highly effective.
People have welcomed conversational AI technologies into our homes, workplaces, and institutions where we interact with them on a daily basis. The proliferation of digital assistants in a multitude of embodiments (e.g., speakers, displays, avatars, robots) in human environments over extended periods of time provides us with new ways to investigate, develop and assess the design of personified AIs that emotionally engage and support people to promote human flourishing across a wide range of applications and usage contexts. In this talk, I highlight a number research projects where we are developing, fielding, and assessing social robots in homes, schools, and living communities of older adults. We explore different embodiments and develop adaptive algorithmic capabilities for our robots to sustain interpersonal engagement and personalize to people's needs to support novel interventions in education, social engagement, and emotional wellness. In addition to evaluating the impact of these capabilities and features on improving learning, sustaining engagement, nudging behavior, and shifting attitudes --- we are also examining the nature of the relationship that people form with these personified AI technologies and how it contributes to these impacts. We conclude by reflecting on the ethical and responsible design of intelligent technologies that emotionally engage and build relationships with people.
Few EC technologies have gone from universities to commercial success. Goodman will describe the SHERPA algorithm, part of the HEEDS design exploration framework, and how Red Cedar Technology, which he co-founded, eventually succeeded. Beginning 20 years ago, SHERPA used a self-adapting ensemble of EC methods (GA, ES, DE, SA, etc.) in each run, requiring no choice of optimization methods or parameters by the engineering user. It is a best-selling engineering design optimizer, still built around the original code developed in 1999-2010, although current owner Siemens now has 20+ developers on HEEDS and SHERPA. Goodman will then turn to a problem outside SHERPA's scope, addressed with unpublished parallel EC methods. NASA provided a futuristic challenge problem to teams of DARPA awardees, to develop ways to optimize the distribution of a set of solid propellant types (eventually to be 3D-printed) in a rocket. Goodman will describe the modeling of the rocket and several problem-specific EC methods used to find feasible solutions to this problem with a design space of 10300--500 and over 700 constraints.
A major challenge of contemporary statistical inference is the large-scale limit, where one wants to discover the values of many hidden parameters, using large amount of data. In recent years, ideas from statistical physics of disordered systems have helped to develop new algorithms for important inference problems, ranging from community detection to compressed sensing, machine learning (notably neural networks), tomography and generalized linear regression. The talk will review these developments and explain how they can be used, to develop new types of algorithms and identify phase transitions.
Evolutionary computation (EC) algorithms involve a careful collaborative and iterative update of a population of solutions to reach near a desired target. In a single-objective search and optimization problem, the respective optimal solution is often the single target. In a multi-criterion optimization problem, the target is a set of Pareto-optimal solutions. Although EC field started with solving single-objective problems, EC researchers soon realized that they were ideal for finding a well-diversed set of multiple Pareto-optimal solutions simultaneously for multi-criterion optimization problems, thereby making a clear niche of EC algorithms compared to their point-based classical counterparts. In this keynote talk, we provide a brief chronology of events on the evolutionary multi-criterion optimization (EMO) field in the past almost three decades, key challenges it faced, and key events and publications which pushed the field forward. We shall also provide a brief account of the current activities and a few pertinent future areas of research and applications.
