[Paper Review] A Super-Polynomial Lower Bound for the Parity Game Strategy Improvement Algorithm as We Know it

Oliver Friedmann|ArXiv.org|Jan 18, 2009

Artificial Intelligence in Games11 references9 citations

TL;DR

This paper presents a family of parity games with quadratic edge count that force the Vöge-Jurdziñski strategy improvement algorithm to perform exponentially many iterations, proving it requires super-polynomial time in the worst case. The construction demonstrates that even locally optimizing improvement policies cannot achieve polynomial-time performance on these instances, resolving a long-standing open question about the algorithm's complexity.

ABSTRACT

This paper presents a new lower bound for the discrete strategy improvement algorithm for solving parity games due to Voege and Jurdziski. First, we informally show which structures are difficult to solve for the algorithm. Second, we outline a family of games of quadratic size on which the algorithm requires exponentially many strategy iterations, answering in the negative the long-standing question whether this algorithm runs in polynomial time. Additionally we note that the same family of games can be used to prove a similar result w.r.t. the strategy improvement variant by Schewe.

Motivation & Objective

• To resolve the long-standing open question of whether the Vöge-Jurdziñski strategy improvement algorithm runs in polynomial time.
• To construct a family of parity games that force the algorithm into exponentially many strategy iterations.
• To demonstrate that the algorithm's worst-case complexity is super-polynomial, despite its strong theoretical appeal.
• To extend the result to Schewe's globally optimizing variant using the same game family.
• To challenge the assumption that strategy improvement algorithms are viable candidates for polynomial-time parity game solving.

Proposed method

• Constructs a family of parity games Gn with O(n) nodes and O(n^2) edges, designed to resist efficient strategy improvement.
• Defines a complex strategy iteration process with nested state transitions across multiple strategy subcomponents.
• Uses a locally optimizing improvement policy that updates strategies based on local gain in value function.
• Employs inductive reasoning on the game structure to prove that each iteration only makes minimal progress toward the optimal strategy.
• Analyzes the sequence of strategy states to show that the algorithm cycles through 13×2^n - 9 distinct strategies before convergence.
• Validates the theoretical findings via empirical benchmarking on the PGSOLVER platform, confirming exponential runtime.

Experimental results

Research questions

• RQ1 Can the Vöge-Jurdziñski strategy improvement algorithm solve all parity games in polynomial time?
• RQ2 Are there specific game structures that force the algorithm to perform exponentially many iterations?
• RQ3 Does the same lower bound apply to Schewe's globally optimizing variant of strategy improvement?
• RQ4 Can preprocessing techniques or hybrid solvers mitigate the exponential behavior on these worst-case instances?
• RQ5 Is strategy improvement fundamentally incapable of achieving polynomial-time performance due to structural limitations in strategy update mechanisms?

Key findings

• The Vöge-Jurdziñski strategy improvement algorithm requires exactly 13 × 2^n - 9 iterations to solve the constructed game family Gn.
• The family of games Gn has O(n) nodes and O(n^2) edges, making it quadratic in size relative to the number of nodes.
• The algorithm's runtime is super-polynomial in the worst case, as each iteration takes polynomial time and the number of iterations is exponential in n.
• The same exponential lower bound applies to Schewe's globally optimizing strategy improvement variant on the same game family.
• Empirical benchmarks on the PGSOLVER platform confirm exponential runtime behavior on these instances.
• The result shows that strategy improvement algorithms cannot be made polynomial-time using only local or global improvement policies, due to structural limitations in how profitable substructures are updated.