Algorithms for partially observable markov decision processes Dissertation Thesis


Author(s): Chmelík, Martin
Advisor(s): Chatterjee, Krishnendu
Title: Algorithms for partially observable markov decision processes
Affiliation IST Austria
Abstract: We study partially observable Markov decision processes (POMDPs) with objectives used in verification and artificial intelligence. The qualitative analysis problem given a POMDP and an objective asks whether there is a strategy (policy) to ensure that the objective is satisfied almost surely (with probability 1), resp. with positive probability (with probability greater than 0). For POMDPs with limit-average payoff, where a reward value in the interval [0,1] is associated to every transition, and the payoff of an infinite path is the long-run average of the rewards, we consider two types of path constraints: (i) a quantitative limit-average constraint defines the set of paths where the payoff is at least a given threshold L1 = 1. Our main results for qualitative limit-average constraint under almost-sure winning are as follows: (i) the problem of deciding the existence of a finite-memory controller is EXPTIME-complete; and (ii) the problem of deciding the existence of an infinite-memory controller is undecidable. For quantitative limit-average constraints we show that the problem of deciding the existence of a finite-memory controller is undecidable. We present a prototype implementation of our EXPTIME algorithm. For POMDPs with w-regular conditions specified as parity objectives, while the qualitative analysis problems are known to be undecidable even for very special case of parity objectives, we establish decidability (with optimal complexity) of the qualitative analysis problems for POMDPs with parity objectives under finite-memory strategies. We establish optimal (exponential) memory bounds and EXPTIME-completeness of the qualitative analysis problems under finite-memory strategies for POMDPs with parity objectives. Based on our theoretical algorithms we also present a practical approach, where we design heuristics to deal with the exponential complexity, and have applied our implementation on a number of well-known POMDP examples for robotics applications. For POMDPs with a set of target states and an integer cost associated with every transition, we study the optimization objective that asks to minimize the expected total cost of reaching a state in the target set, while ensuring that the target set is reached almost surely. We show that for general integer costs approximating the optimal cost is undecidable. For positive costs, our results are as follows: (i) we establish matching lower and upper bounds for the optimal cost, both double and exponential in the POMDP state space size; (ii) we show that the problem of approximating the optimal cost is decidable and present approximation algorithms that extend existing algorithms for POMDPs with finite-horizon objectives. We show experimentally that it performs well in many examples of interest. We study more deeply the problem of almost-sure reachability, where given a set of target states, the question is to decide whether there is a strategy to ensure that the target set is reached almost surely. While in general the problem EXPTIME-complete, in many practical cases strategies with a small amount of memory suffice. Moreover, the existing solution to the problem is explicit, which first requires to construct explicitly an exponential reduction to a belief-support MDP. We first study the existence of observation-stationary strategies, which is NP-complete, and then small-memory strategies. We present a symbolic algorithm by an efficient encoding to SAT and using a SAT solver for the problem. We report experimental results demonstrating the scalability of our symbolic (SAT-based) approach. Decentralized POMDPs (DEC-POMDPs) extend POMDPs to a multi-agent setting, where several agents operate in an uncertain environment independently to achieve a joint objective. In this work we consider Goal DEC-POMDPs, where given a set of target states, the objective is to ensure that the target set is reached with minimal cost. We consider the indefinite-horizon (infinite-horizon with either discounted-sum, or undiscounted-sum, where absorbing goal states have zero-cost) problem. We present a new and novel method to solve the problem that extends methods for finite-horizon DEC-POMDPs and the real-time dynamic programming approach for POMDPs. We present experimental results on several examples, and show that our approach presents promising results. In the end we present a short summary of a few other results related to verification of MDPs and POMDPs.
Publication Title: IST Dissertation
Degree Granting Institution: IST Austria  
Degree: PhD
Degree Date: 2016-02-01
Start Page: 1
Total Pages: 232
Open access: no
IST Austria Authors
  1. Martin Chmelik
    23 Chmelik
Related IST Austria Work