Approximation of Bound Functions in Algorithms for Solving Stochastic Games
Aproximace konvexních funkcí v algoritmech pro řešení stochastických her
Authors
Supervisors
Reviewers
Editors
Other contributors
Journal Title
Journal ISSN
Volume Title
Publisher
České vysoké učení technické v Praze
Czech Technical University in Prague
Czech Technical University in Prague
Date
Files
Abstract
V této práci se soustředíme na aproximaci konvexních funkcí v Heuristic Search Value Iteration algoritmu pro řešení Jednostranně Částečně Pozorovatelných Stochastických Her. Jedná se o dynamické hry, kde první hráč má neúplnou informaci o hře, zatímco druhý hráč má informaci úplnou. Konvexní funkce tvoří odhady tzv. value funkce celé hry. Dolní odhad je tvořen pomocí horní obálky lineárních funkcí, zatímco horní odhad je tvořen jako dolní konvexní obálka množiny bodů. V práci se zaměřujeme pouze na aproximaci horního odhadu převážně pomocí Aproximativního Convex Hull algoritmu. Ukazujeme, že aproximace horního odhadu je problematická a že pro lepší výsledky je zapotřebí se zaměřit také na aproximaci dolního odhadu.
In this thesis, we focus on the approximation of the bound functions in the Heuristic Search Value Iteration (HSVI) algorithm for One-Sided Partially Observable Stochastic Games (OS-POSG). These are dynamic games with infinite horizon where only one player has imperfect information, and the opponent has full information. The bound functions approximate the value function of the game. The lower bound is represented as an upper envelope of linear functions, while the upper bound is represented as a lower convex envelope of a set of points. We focus only on the approximation of the upper bound mainly by using the Approximate Convex Hull algorithm. We show that the approximation of the upper bound is problematic and that for better results, it is necessary to focus on the approximation of the lower bound function as well.
In this thesis, we focus on the approximation of the bound functions in the Heuristic Search Value Iteration (HSVI) algorithm for One-Sided Partially Observable Stochastic Games (OS-POSG). These are dynamic games with infinite horizon where only one player has imperfect information, and the opponent has full information. The bound functions approximate the value function of the game. The lower bound is represented as an upper envelope of linear functions, while the upper bound is represented as a lower convex envelope of a set of points. We focus only on the approximation of the upper bound mainly by using the Approximate Convex Hull algorithm. We show that the approximation of the upper bound is problematic and that for better results, it is necessary to focus on the approximation of the lower bound function as well.
Description
Keywords
Teorie her, Jednostranně Částečně Pozorovatelné Stochastické Hry, Markovovy Rozhodovací Procesy, Částečně Pozorovatelné Markovovy Rozhodovací Procesy, Heuristic Search Value Iteration algoritmus, Konvexní obal, Aproximativní konvexní obal, Game Theory, One-Sided Partially Observable Stochastic Games, Markov Decision Processes, Partially Observable Markov Decision Processes, Heuristic Search Value Iteration algorithm, Convex hull, Approximate convex hull