Saturated Path-Constrained MDP: Planning under Uncertainty and Deterministic Model-Checking Constraints

Authors: Jonathan Sprauel, Andrey Kolobov, Florent Teichteil-Königsbuch

AAAI 2014 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We propose a dynamic programming-based algorithm for finding such policies, and empirically demonstrate this algorithm to be orders of magnitude faster than its next-best alternative. ... The objectives of our experiments are two-fold: (1) to compare SPC VI with the algorithm for PC MDPs, PC MDP-ILP (Teichteil-Konigsbuch 2012), in terms of efficiency, and (2) to validate SPC MDPs as an efficient modeling tool. The experiments were run with 5.8 GB of RAM on a 2.80GHz CPU.
Researcher Affiliation Collaboration Jonathan Sprauel and Florent Teichteil-K onigsbuch {jonathan.sprauel, florent.teichteil}@onera.fr ONERA The French Aerospace Lab 2 Avenue Edouard-Belin, F-31055 Toulouse, France Andrey Kolobov akolobov@microsoft.com Microsoft Research Redmond, WA-98052, USA
Pseudocode Yes Algorithm 1: SPC MDP Value Iteration, Algorithm 2: explore(S, Tip) function, Algorithm 3: update Reachability(S, ξi, θ) function
Open Source Code No The paper does not provide a direct link or an explicit statement about the availability of its source code.
Open Datasets No We randomly generated instances with different grid dimensions (between 10x10 and 100x100, with a total number of states between 200 and 160 000), time parameter values, and numbers of zones of each type (between 1 and 5 per type). ... We tested instances having from 1 computer (1246 states) to 3 computers (1 014 013 states). The paper describes generating its own instances rather than using or providing concrete access to a publicly available dataset.
Dataset Splits No The paper does not explicitly describe training, validation, and test dataset splits.
Hardware Specification Yes The experiments were run with 5.8 GB of RAM on a 2.80GHz CPU.
Software Dependencies No The paper mentions using the "PPDDL language" but does not specify any software names with version numbers or other key software dependencies.
Experiment Setup Yes In all experiments, the SPC MDP discount factor was γ = 0.9. ... For ϵoptimality we set ϵ = 0.1, with a corresponding ω = 0.001, since we fixed the penalty of a fire occurrence to -1. ... Since the reward function ranges from -50 to 50, we chose a parameter ϵ 10 for the ϵoptimality; the corresponding ω parameter is 0.001.