LGOCMLNov 28, 2019

Analysis of Lower Bounds for Simple Policy Iteration

arXiv:1911.12842v13 citations
Originality Incremental advance
AI Analysis

This work addresses a theoretical gap in understanding the computational complexity of policy iteration algorithms for researchers in reinforcement learning and optimization, though it is incremental as it builds directly on prior results.

The paper tackles the problem of generalizing exponential lower bounds for Simple Policy Iteration (SPI) in Markov Decision Processes (MDPs), proving a novel exponential lower bound of O((3+k)2^(N/2-3)) for N-state, k-action MDPs.

Policy iteration is a family of algorithms that are used to find an optimal policy for a given Markov Decision Problem (MDP). Simple Policy iteration (SPI) is a type of policy iteration where the strategy is to change the policy at exactly one improvable state at every step. Melekopoglou and Condon [1990] showed an exponential lower bound on the number of iterations taken by SPI for a 2 action MDP. The results have not been generalized to $k-$action MDP since. In this paper, we revisit the algorithm and the analysis done by Melekopoglou and Condon. We generalize the previous result and prove a novel exponential lower bound on the number of iterations taken by policy iteration for $N-$state, $k-$action MDPs. We construct a family of MDPs and give an index-based switching rule that yields a strong lower bound of $\mathcal{O}\big((3+k)2^{N/2-3}\big)$.

Foundations

The foundational work for this paper's niche, ranked by how specifically the neighbourhood builds on it — not by global fame.

Your Notes