2020
Reinforcement Learning of Risk-Constrained Policies in Markov Decision Processes
BRÁZDIL, Tomáš; Krishnendu CHATTERJEE; Petr NOVOTNÝ a Jiří VAHALAZákladní údaje
Originální název
Reinforcement Learning of Risk-Constrained Policies in Markov Decision Processes
Autoři
BRÁZDIL, Tomáš; Krishnendu CHATTERJEE; Petr NOVOTNÝ a Jiří VAHALA
Vydání
Palo Alto, California, USA, The Thirty-Fourth AAAI Conference on Artificial Intelligence, AAAI 2020, od s. 9794-9801, 8 s. 2020
Nakladatel
AAAI Press
Další údaje
Jazyk
angličtina
Typ výsledku
Stať ve sborníku
Obor
10200 1.2 Computer and information sciences
Stát vydavatele
Spojené státy
Utajení
není předmětem státního či obchodního tajemství
Forma vydání
elektronická verze "online"
Odkazy
Označené pro přenos do RIV
Ano
Kód RIV
RIV/00216224:14330/20:00114279
Organizační jednotka
Fakulta informatiky
ISBN
978-1-57735-823-7
UT WoS
EID Scopus
Klíčová slova anglicky
reinforcement learning; Markov decision processes; Monte Carlo tree search; risk aversion
Štítky
Příznaky
Mezinárodní význam, Recenzováno
Změněno: 15. 5. 2024 01:27, RNDr. Pavel Šmerk, Ph.D.
Anotace
V originále
Markov decision processes (MDPs) are the defacto framework for sequential decision making in the presence of stochastic uncertainty. A classical optimization criterion for MDPs is to maximize the expected discounted-sum payoff, which ignores low probability catastrophic events with highly negative impact on the system. On the other hand, risk-averse policies require the probability of undesirable events to be below a given threshold, but they do not account for optimization of the expected payoff. We consider MDPs with discounted-sum payoff with failure states which represent catastrophic outcomes. The objective of risk-constrained planning is to maximize the expected discounted-sum payoff among risk-averse policies that ensure the probability to encounter a failure state is below a desired threshold. Our main contribution is an efficient risk-constrained planning algorithm that combines UCT-like search with a predictor learned through interaction with the MDP (in the style of AlphaZero) and with a risk-constrained action selection via linear programming. We demonstrate the effectiveness of our approach with experiments on classical MDPs from the literature, including benchmarks with an order of 10^6 states.
Návaznosti
| GA18-11193S, projekt VaV |
| ||
| GA19-15134Y, interní kód MU |
| ||
| GJ19-15134Y, projekt VaV |
| ||
| MUNI/A/1050/2019, interní kód MU |
| ||
| MUNI/G/0739/2017, interní kód MU |
|