Adaptive Choice Set Construction for Complex Decision Making
Abstract
Adaptive Choice Set Construction for Complex Decision MakingCurrently we have limited understanding of how to define choice sets forautonomous systems. Many cutting-edge solutions, such as the ~proposal networks~ implemented by AlphaGo and AlphaGo Zero, come with notable limitations. First, they use deep neural networks that deliver impressive performance but are ~black boxes~, affording no insight into the underlying mechanisms. Second, they require massive training sets including, in some cases, extensive exposure to human expert play. Third, they provide a solution to one specific task (e.g.,playing Go), but do not suitably generalize to the general problem of choice set construction that arises across diverse tasks. A major barrier to progress in the field is the lack of suitably general algorithms for choice set construction.Our proposal aims squarely at this barrier. Our goal is to define computationally precise and general algorithms for choice set construction. We want these algorithms to be ~plug-andplay~ for other researchers, whatever decision problem their autonomous system is designed to solve. Thus, our approach is not to build a full cognitive architecture for a single task (e.g., onesingle, novel system that proceeds from raw sense data to motor output). Rather, our approach is to build algorithmic ~modules~ for the specific problem of choice set construction, and to make these general enough that they could be adopted into a variety of different specific autonomous systems implemented by other researchers. In other words, we propose a program of basic researchthat delivers a final product ready to interface with diverse kinds of implementations. Our approach to this problem is to ~reverse engineer~ human choice set construction. We know that humans make remarkably adaptive decisions in very complex environments. But, we don~t know how they accomplish this. Nearly all experimental research on human decision-makingartificially solves the ~choice set construction~ process by presenting humans with exactly two choices. In contrast, we propose new experimental research where humans face many choices. We then fit computational cognitive models to the human data, refining these models in order to determine how, at an algorithmic level, humans construct adaptive choice sets. In summary: Wecan build better and more general algorithms for choice set construction by understanding how humans solve this problem. Our research will accomplish this goal.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jan 23, 2019
- Source ID
- N000141912025
Entities
People
- Fiery Cushman
Organizations
- Office of Naval Research
- President and Fellows of Harvard College
- United States Navy