[Paper Review] On-the-fly Fast Mean-Field Model-Checking: Extended Version

Diego Latella, Michele Loreti+1 more|arXiv (Cornell University)|Dec 12, 2013

Formal Methods in Verification11 references3 citations

TL;DR

This paper presents an on-the-fly, scalable model-checking procedure that combines local probabilistic model checking with discrete-time mean-field approximation to verify bounded PCTL properties of selected individuals in large populations of independent, interacting objects. The method achieves asymptotic correctness and is insensitive to population size, enabling fast verification via a prototype tool, FlyFast, as demonstrated on a computer epidemics benchmark.

ABSTRACT

A novel, scalable, on-the-fly model-checking procedure is presented to verify bounded PCTL properties of selected individuals in the context of very large systems of independent interacting objects. The proposed procedure combines on-the-fly model checking techniques with deterministic mean-field approximation in discrete time. The asymptotic correctness of the procedure is shown and some results of the application of a prototype implementation of the FlyFast model-checker are presented.

Motivation & Objective

• To address the scalability limitations of traditional model checking in large systems with many independent, interacting objects.
• To develop a fast, on-the-fly model-checking procedure that remains efficient regardless of population size.
• To integrate deterministic mean-field approximation with local, on-the-fly probabilistic model checking for bounded PCTL formulas.
• To enable verification of individual object properties in the context of global system behavior without exploring the full state space.
• To provide a parametric framework supporting both exact probabilistic semantics and mean-field approximations.

Proposed method

• Uses a discrete-time, time-synchronous model where each object evolves via a finite-state Markov chain influenced by the global occupancy measure.
• Applies mean-field approximation to represent the system’s global behavior as a deterministic difference equation, capturing the fraction of objects in each local state.
• Integrates on-the-fly model checking with the mean-field approximation by recursively evaluating PCTL formulas on a hybrid state space of individual object states and global mean-field measures.
• Employs a recursive, top-down traversal of the formula structure, computing path probabilities dynamically while leveraging time-varying truth values based on the current mean-field state.
• Implements the algorithm in FlyFast, a prototype model-checker that instantiates the procedure using the limit DTMC (discrete-time Markov chain) derived from the mean-field approximation.
• Uses safety checks on probability thresholds to ensure correctness during recursive formula evaluation, with a weaker condition sufficient for practical use.

Experimental results

Research questions

• RQ1 Can on-the-fly model checking be effectively combined with mean-field approximation to verify bounded PCTL properties in large-scale systems?
• RQ2 Does the proposed method remain scalable and accurate as the population size increases?
• RQ3 How can time-varying truth values of formulas be handled efficiently in a dynamic, mean-field context?
• RQ4 What is the relationship between the exact probabilistic semantics and the mean-field approximation in this on-the-fly setting?
• RQ5 Can the method be extended to heterogeneous systems and more expressive logics while preserving correctness and efficiency?

Key findings

• The proposed on-the-fly mean-field model-checking procedure is asymptotically correct under the assumptions of Theorem 4.1 in [25], ensuring convergence to the true probability as population size grows.
• The FlyFast prototype performs verification in less than one second on a large-population computer epidemics model, with performance independent of population size.
• The method correctly handles time-varying truth values of PCTL formulas by dynamically updating them based on the current mean-field state at each time step.
• The algorithm supports nested path formulae and probabilistic operators such as P>0.3(φ U≤k ψ), with verified correctness through safety checks on probability thresholds.
• The approach enables scalable verification of individual object properties in systems with thousands or more objects, without explicit state space enumeration.
• The results demonstrate that mean-field approximation can be effectively integrated into on-the-fly model checking, preserving logical expressiveness and computational efficiency.