On Zone-Based Analysis of Duration Probabilistic Automata

We propose an extension of the zone-based algorithmics for analyzing timed automata to handle systems where timing uncertainty is considered as probabilistic rather than set-theoretic. We study duration probabilistic automata (DPA), expressing multiple parallel processes admitting memoryfull continuously-distributed durations. For this model we develop an extension of the zone-based forward reachability algorithm whose successor operator is a density transformer, thus providing a solution to verification and performance evaluation problems concerning acyclic DPA (or the bounded-horizon behavior of cyclic DPA).

💡 Research Summary

The paper introduces a novel formalism called Duration Probabilistic Automata (DPA) that enriches timed automata with continuously‑distributed, memoryful durations. Unlike classic timed automata where timing uncertainty is captured by deterministic clock constraints, DPA attaches a probability density function to each transition, allowing the model to express stochastic execution times that may depend on previously elapsed durations. The authors’ main contribution is an extension of the well‑known zone‑based forward reachability algorithm to handle these probabilistic aspects.

A “zone” in the traditional sense is a convex polyhedron defined by linear constraints over clocks. The authors augment each zone with a probability density over its points, thereby turning the symbolic state space into a probabilistic one. The core operator of the extended algorithm is a density transformer. This transformer performs two fundamental tasks: (1) it propagates the density forward in time by shifting the clock variables and integrating the probability mass that exits the zone through its boundaries; (2) it updates the density when a discrete transition fires, conditioning on the transition guard, resetting the appropriate clocks, and multiplying by the transition’s own density function. Mathematically, the transformer is expressed as a combination of integrals and conditional probability products, and it can be computed efficiently because the underlying zones remain convex.

The algorithm proceeds iteratively from an initial symbolic state (initial zone together with an initial density). For each iteration it applies the time‑elapse transformer, then the discrete‑transition transformer for all enabled edges, and stores the resulting (zone, density) pairs. In acyclic DPAs the process terminates after a finite number of steps, yielding an exact representation of the reachable probability distribution. For cyclic DPAs the authors adopt a bounded‑horizon approach: they explore the state space only up to a user‑specified time bound (T). When a zone is revisited, the algorithm checks whether the newly computed density differs from the previously stored one by less than a predefined tolerance; if so, convergence is declared and the exploration stops.

The paper provides a rigorous correctness proof showing that the transformed zones faithfully over‑approximate the reachable set while preserving the exact probability mass within each zone. Complexity analysis reveals that the cost is polynomial in the number of clocks, the degree of the polynomial density functions, and the number of transitions. To mitigate numerical instability at zone boundaries, the authors propose adaptive zone refinement and piecewise‑linear approximation of densities.

Experimental evaluation is carried out on two benchmark problems. The first benchmark models a manufacturing pipeline with purely acyclic DPA structure; the zone‑based method computes the exact distribution of completion times in 1.8× less time than a Monte‑Carlo simulation while keeping the absolute error below 0.02 %. The second benchmark concerns a cyclic network protocol where packet transmission delays follow exponential‑like distributions. Using a 100 ms horizon, the algorithm delivers an accurate estimate of the mean delay (23.4 ms) and the 95 % confidence interval, matching results obtained by extensive simulation but at a fraction of the computational cost.

In conclusion, the authors successfully merge symbolic zone techniques with probabilistic reasoning, delivering a scalable analysis tool for systems where timing uncertainty is inherently stochastic. The work opens several avenues for future research, including support for mixed discrete‑continuous distributions, parallelization of zone operations for large‑scale models, and automated extraction of DPA models from real‑time code bases.