Steve Kroon PLEASED: Planning, Learning, and Search for - PowerPoint PPT Presentation

Steve Kroon PLEASED: Planning, Learning, and Search for Decision-making. http://www.cs.sun.ac.za/~kroon/decision.html Maties Machine Learning: 25 October 2019

“This group considers almost any aspect of the general decision-making problem, including sequential decision-making under uncertainty. Major sub-problems we consider are planning, machine learning, and search algorithms. Our approach is grounded in probability theory and game theory for managing uncertainty and multi-agent systems.” Images: https://mimiandeunice.com/wp-content/uploads/2011/08/ME_447_Decisions-640x199.png

Single agent, single decision Principled - grounded in: ● Probability theory ● Decision theory ● Game theory Typically requires: ● A model (perhaps >1) ● Data ● Payoff function Good foundation: Bayesian decision theory Images:https://www.azimuthproject.org/azimuth/files/BayesianSDT-bigpic.png

Extensions ● Sequential decision making ○ Search ○ Planning ○ Bayes filter ○ Reinforcement learning ● Multi-agent settings ○ Adversarial ○ Collaborative ● Tractable inference/decision making ○ Inference approaches ○ Search techniques ○ Choice of approximations Images:https://www.azimuthproject.org/azimuth/files/BayesianSDT-bigpic.png

Group members

Learning theory * *understanding relationships between and properties of machine learning/statistical models and approaches to fitting them

Search/Planning

Bayesian analysis Images: https://miro.medium.com/max/1002/1*hblsrFOWViHS43l5YpUXeQ.png

Latent variable models/ variational inference Images:https://miro.medium.com/max/800/1*pZo_IcxW1GVuH2vQKdoIMQ.jpeg

Process monitoring and fault detection/diagnosis Images:https://www.ericsson.com/49d220/assets/global/qbank/2019/06/13/architecture-50-109173resize436234crop00436234autoorientquality90stripbackground23ffffffextensionjpgid8.jpg

Specific interests Learning Theory* (mostly NNs) Common elements: Search/Planning (mostly MCTS) Regularizing effects of model, ● inference, and optimization Bayesian analysis Tractable inference/search ● Dynamical systems ● Latent variable models and variational inference Process monitoring, fault detection and diagnosis *understanding relationships between and properties of machine learning/statistical models and approaches to fitting them

THANK YOU - QUESTIONS?