Control of Covid-19 using Agent-based modelling with Reinforcement learning.
Authors
Polymenidis, Stergios ; Wawrzyniak, Bartosz
Term
4. term
Education
Publication year
2021
Submitted on
2021-06-03
Pages
96
Abstract
Smitsomme sygdomme har længe udfordret samfund, og Covid-19 er et aktuelt eksempel: sygdommen er sjældent dødelig, men spreder sig hurtigt og kan komme ud af kontrol. Regeringer står derfor over for opgaven at bremse smitten uden at overbelaste sundhedsvæsenet. I dette projekt undersøger vi, om forstærkningslæring (reinforcement learning) kan finde en optimal rækkefølge af tiltag, som holder antallet af indlagte og patienter med behov for respirator inden for fastsatte grænser. Vi bygger en agentbaseret model, der simulerer et samfund og virusspredning gennem interaktioner mellem individer. Med udgangspunkt i virkelige myndighedsreaktioner opstiller vi en liste over politiske tiltag, som i simuleringen fremkalder reaktioner, der ligner dem i den virkelige verden. Vi anvender en model-fri værdi-iteration, en form for forstærkningslæring, der gennem prøve-og-fejl i simuleringen lærer, hvilke beslutninger der giver de bedste resultater, til at finde en sekvens af tiltag, der begrænser smitten og holder belastningen på hospitaler under tærsklerne. Vi konstruerer tre modeller med forskellig kompleksitet for at afprøve metoden. Vi simulerer to af modellerne, og resultaterne viser, at algoritmen kan finde den ønskede sekvens af tiltag.
Infectious diseases have long challenged societies, and Covid-19 is a current example: the disease is rarely fatal but spreads quickly and can get out of control. Governments therefore face the task of slowing transmission without overwhelming healthcare. In this project, we test whether reinforcement learning can identify an optimal sequence of public policies that keeps hospitalizations and severe cases requiring ventilators within set thresholds. We build an agent-based model that simulates a society and how the virus spreads through interactions between individuals. Based on real government responses, we define a set of policy options that elicit reactions in the simulation similar to those observed in the real world. We apply a model-free value-iteration reinforcement learning algorithm—an approach that learns via trial and error in the simulation which choices lead to the best outcomes—to discover policy sequences that control spread and keep demand on hospitals within limits. We design three models with different levels of complexity to evaluate the approach. We simulate two of these models, and the results show that the algorithm can find the desired sequence of policies.
[This summary has been rewritten with the help of AI based on the project's original abstract]
Documents