 |
Center for Nonlinear Studies |
Probability Collectives is a set of optimization methods that arose from recognizing connections between game theory, machine learning, statistical physics, and randomized optimization algorithms. These methods target large scale optimization problems involving categorical variables, and convert them to continuous problems over smaller spaces. This is done by considering expectations over appropriate probability distributions. By suitable specification of these distributions, various aspects of particular problems can be incorporated, like the structure of dependencies, decentralization, prior knowledge, etc. In the end, these methods yield probability distributions over solutions. They are easily implemented in an adaptive, decentralized framework, so that the resulting solutions represent good decentralized, adaptive, stochastic strategies, rather than fixed optimal solutions. The talk will introduce the algorithm, discuss some implementation details, and demonstrate some examples.