Modernized Decision Making

Evolutionary Surrogate-assisted Prescription (ESP)

Introduction

How can we make good decisions in business, engineering design, science, education, and indeed in life in general? Much of good decisions are based on experience: recalling what decisions were made in similar situations in the past, how well they worked out, and modifying them to achieve good outcomes in current situation. But there is also an element of trial and error as well. The situation maybe different from anything that was seen before, or it may be desirable to do better than in the past.

Evolutionary Surrogate-assisted Prescription (ESP) is a machine learning technology that makes it possible to come up with good decision strategies automatically. The idea is to use historical data to build a predictive surrogate model, and population based search (i.e. evolutionary computation) to discover good decision strategies. Each strategy is evaluated with the surrogate instead of the real world, so that millions of strategies can be tested before they are deployed in the real world (where mistakes may be costly).

This site gives a high-level summary of ESP, and presents two examples: optimizing behavior in sequential decision tasks, and optimizing non-pharmaceutical interventions in the COVID-19 pandemic. Each example includes links to papers, video presentations, animations, and also an interactive demo for Covid-19 interventions.

I. The ESP Framework for Optimizing Decision Making

The general ESP framework consists of a Predictor (a surrogate model) and a Prescriptor (the decision policy), in an outer loop that allows the system to adapt to a changing environment. It can be applied to static or sequential decision making tasks, as is described in the two examples below.

II. Sequential Decision Making: Data-Efficient RL with ESP

In sequential decision-making tasks, the Predictor and Prescriptor receive a sequence of inputs and make a sequence of decisions, with performance evaluated in the end. Compared to standard reinforcement learning, the process is data-efficient and reliable, as illustrated in various visualization and game-playing domains.

III. Augmenting Human Decision Making: Optimizing COVID-19 Interventions

The COVID-19 pandemic presents an important challenge: What non-pharmaceutical interventions can be taken to contain the spread while impacting the economy as little as possible? Based on available data, ESP can be used to make such recomendations and evaluate how well they would work.

Contributing Researchers

Babak Hodjat: Team Lead; PhD Kyushu University 2003
Risto Miikkulainen: Research Lead; Professor of Computer Science at UT Austin; PhD UCLA 1990
Hormoz Shahrzad: Senior Research Scientist
Olivier Francon: Senior Research Engineer
Xin Qiu: Senior Research Scientist; PhD National University of Singapore 2016
Elliot Meyerson: Research Scientist; PhD UT Austin 2018
Santiago Gonzalez: Research Scientist; PhD UT Austin 2020 (expected)