The Hitchiker’s Guide to Stochastic Games
This is the web-page for our work on solving stochastic games in a sound and efficient way through value iteration and partial exploration. In case you have any questions, please contact me or Maxi (click to write an email).
Markov Chain Gang - Spurious Fixpoints Suck (Verse 1) Yo, gather 'round, let me tell you 'bout what we found, We're doing value iteration in a way that's sound In stochastic games on graphs they're played round by round There we complement VI with converging bounds Arbitrary imprecision, yea, that ain't no more We unify all the papers that came before They were all restricted to particular objectives Or to a single player, but we're not that selective So you ask: Why should I care? Does this solve any problems that I share? Let me give you some words that are key: Probabilistic model checking; game theory If you wanna model check your stochastic game But it's too large – scalability's the name of the aim value iteration scales so well it's insane and we fixed it – that's our claim to fame (Hook 2x) Bellman updates - can get stuck Spurious fixpoints - man they suck Can you stop? - I don't know Why? - cause my bounds they are not close (Verse 2) Like I said, Bellmann updates they can linger At a fixpoint that is wrong, for example bigger Than it should, yea that's sour, baby, just like vinegar Here's how to solve objectives that are fixpoint-linear Step one: strategy recommender is what you need We prove that classical VI allows us to proceed Step two: use the recommender to find SECs Simple end components - get your mind outta the gutter! Step three – now we really start flexin: observe that in a cycle you can only stay there or exit So when you set your estimate to match the best of the two You can force yourself to get out of the Bellman stuck blues Our approach is objective-independent it's novel, unifying, almost transcendant We call our procedures deflate or inflate And their based on asking: should I go or should I stay (Bridge) This comes from IST in Klosterneuburg But for real science I got that international support I got my bro in Lancester, the Leipzig resort And from Masaryk in Brno came the initial thought All of us met at the TU of Munich And our first draft of this was just cryptic and runic With time we found the unifying solutions VI's stopping - criteria - are our - contribution (Hook 2x) Bellman updates - can get stuck Spurious fixpoints - man they suck Can you stop? - Hey I know Why? - cause I asked should I stay or go (Verse 3) So next time you're in doubt and you wonder what to do With your stochastic system exhibiting a player or two Remember always that guaranteed precision is crucial And use our unified value iteration as the solution Get the gist: check out our paper I'm praying And if you ask: should I be going or staying I tell you: stay, don't you hear the music is playing Listen one more time to that sick hook cause it's slaying (Hook 2x) Bellman updates - can get stuck Spurious fixpoints - man they suck Can you stop? - Hey I know Why? - cause I asked should I stay or go Should I stay or should I go now? Epsilon-close Stop
Theory: Should I Stay or Should I Go?
Our paper “Stopping Criteria for Value Iteration on Stochastic Games with Quantitative Objectives” by Jan Křetínský, Tobias Meggendorfer, and Maximilian Weininger (LICS 2023), presents a sound stopping criterion for value iteration in stochastic games in a unified way. The central slogan is to ask “Should I stay or should I go?”, which neatly identifies and fixes the core convergence problem of (interval) value iteration, namely spurious fixed points.
Click any of the buttons below to access the relevant details:
To cite, you can use the following BibTeX entry (or download it here):
@inproceedings{DBLP:conf/lics/KretinskyMW23,
author = {Jan K{\v{r}}et{\'{\i}}nsk{\'{y}} and
Tobias Meggendorfer and
Maximilian Weininger},
title = {Stopping Criteria for Value Iteration on Stochastic Games with Quantitative
Objectives},
booktitle = {{LICS}},
pages = {1--14},
year = {2023},
doi = {10.1109/LICS56636.2023.10175771}
}
Practice: Partial Exploration Tool 2.0
In our paper “Playing Games with your PET: Extending the Partial Exploration Tool to Stochastic Games” by Tobias Meggendorfer and Maximilian Weininger (CAV 2024), we implement the above algorithms in our tool PET (partial exploration tool). Furthermore, PET also extends partial exploration to stochastic games and comes with many engineering improvements. The paper received a “Distinguished Paper Award” at CAV 2024. The artefact obtained both the available and reusable badge.
Click any of the buttons below to access the relevant details:
Cite information will be added once the paper is published.