Registered Data

[00702] Sequential Decision Making for Optimization, Learning and Search

  • Session Date & Time : 4C (Aug.24, 13:20-15:00)
  • Type : Proposal of Minisymposium
  • Submitted : 2022-12-21
  • Abstract : Key problems such as hyperparameter optimization, model calibration, and inverse/optimal design often involve exploring design spaces to identify desirable designs for one or more objectives of great value and great cost. Intelligently experimenting in this design space is fundamental to gaining valuable, actionable insights in a viable amount of time. In this minisymposium, we will discuss some of the common methodologies for identifying high-performing and optimal designs, including Bayesian and genetic methods, and several exciting applications which motivate the research in this field.
  • Organizer(s) : Michael McCourt
  • Classification : 60G15, 90C26, 62C10
  • Speakers Info :
    • Michael McCourt (SigOpt, an Intel company)
    • Hideaki Imamura (Preferred Networks, Inc)
    • Masahiro Nomura (CyberAgent)
    • Jungtaek Kim (University of Pittsburgh)
  • Contact Person : Michael McCourt (SigOpt, an Intel company)
  • Talks in Minisymposium :
    • [04910] Combinatorial 3D Shape Assembly with Sequential Decision-Making Processes
      • Author(s) :
        • Jungtaek Kim (University of Pittsburgh)
      • Abstract : We require unit primitives, e.g., voxels and points, to create a 3D shape. In particular, if we consider a way to construct a 3D shape with the connectivity of primitives, a problem of 3D shape creation is characterized by sequential and combinatorial properties. By dealing with the sequential and combinatorial properties, we present a method for 3D shape assembly using sequential decision-making processes, i.e., Bayesian optimization and reinforcement learning.
    • [05274] Constraint active search as an alternative to optimization
      • Author(s) :
        • Michael McCourt (SigOpt, an Intel company)
        • Michael McCourt (Unaffiliated)
      • Abstract : Bayesian optimization is a sample efficient method for identifying high performing configurations of a black box function. This strategy is extremely powerful, but it is often a misguided tool for many practical circumstances -- problems with heavy noise, input/output imprecision, many objectives, discrepancy in cost of objective evalution, or a human-in-the-loop defined objective/preference all are situations where optimization may be the wrong strategy. Here, we discuss the shortcomings of optimization and propose an alternate strategy: the search for a satisfactory set of outcomes, as guided by user-defined performance thresholds. We refer to this as Constraint Active Search, and we present our motivating application as well as some theoretical analysis.
    • [05288] Evolution Strategies: Principles and Practical Issues
      • Author(s) :
        • Masahiro Nomura (CyberAgent)
      • Abstract : Evolution strategies (ES) is one of the most powerful frameworks for black-box continuous optimization. This talk will describe the design principles behind the empirical success of ES and the representative methods that have often been employed in science and industry. In addition, key issues that may be encountered when using ES in practice will be discussed.