Accession Number:

ADA591909

Title:

Algorithms for Learning and Decision Making

Descriptive Note:

Final technical rept. 1 Aug 2010-31 Jul 2013

Corporate Author:

MASSACHUSETTS INST OF TECH CAMBRIDGE LAB FOR INFORMATION AND DECISION SYSTEMS

Personal Author(s):

Report Date:

2013-12-09

Pagination or Media Count:

5.0

Abstract:

We have investigated learning algorithms for inference and decision making, by using exact and approximate optimization methods. Most of our research has been in approximate dynamic programmingreinforcement learning methods, with a focus on Markovian Decision Problems with a very large number of states. Much of our work is related to a fundamental algorithm, Q-learning, and related new methods that relate to exact and approximate policy iteration. In particular, we have investigated, convergence issues, error bounds, policy oscillation, exploration-enhanced methods, and issues of decision making in a multi-agent environment. Another research area is large-scale convex optimization methods, with a focus on problems whose cost function involves a sum of a large number of component functions. This includes a unifying framework for polyhedral approximation recently proposed by the principal investigator, incremental gradient and subgradient methods, which are currently at the forefront of algorithmic machine learning research, as well as a new incremental version of the proximal minimization algorithm. We have developed new polyhedral approximation algorithms, including a simplicial decomposition method that applies to large-scale conic programming problems.

Subject Categories:

  • Administration and Management
  • Psychology
  • Numerical Mathematics

Distribution Statement:

APPROVED FOR PUBLIC RELEASE