The study, published in Frontiers in Manufacturing Technology by Konstantinos S. Boulas and Georgios D. Dounias of the University of the Aegean and Chrissoleon T. Papadopoulos of Aristotle University of Thessaloniki, tackles a problem that sits at the heart of manufacturing science. Production lines are the engines of the world economy, and their most valuable performance metric is throughput, the average number of finished parts produced per unit of time by the last machine in the system. Yet even small lines suffer from the notorious state explosion problem: the number of possible system configurations grows so fast that exact symbolic solutions become unwieldy almost immediately.

The researchers modeled short serial production lines as finite Markov chains, stochastic processes in which each machine can be idle, working, or blocked, and the entire system state is described by a string of digits, one per machine. Under the model assumptions, the first station is never starved because it draws from an infinite warehouse of raw parts, the last station is never blocked because finished products flow into unlimited storage, processing times follow negative exponential distributions with mean service rates, and machines are perfectly reliable. The absence of intermediate buffers makes the lines challenging to operate, because blocking and starving phenomena propagate through the system and reduce throughput.

The methodological innovation lies in how the steady-state probabilities are extracted. Instead of solving the linear equilibrium equations directly, the team constructs the state transition diagram as a graph, using a purpose-built algorithm that generates all feasible system states and all transitions between them, each weighted by the mean service rate of the machine that causes the transition. A recently introduced technique from Part I of this research program then assigns symbolic probabilities to the graph anti-arborescences, the reversed spanning trees rooted at each vertex. Each anti-arborescence corresponds to a product of mean service rates, forming a monomial, and the sum of all monomials ending at a given state forms that state’s spanning polynomial. Dividing each state polynomial by the sum of all polynomials across the graph normalizes the probabilities to one, yielding the exact stationary distribution as a ratio of two polynomials.

For the two-station line, the graph is small enough that the anti-arborescences can be counted by hand, and the resulting formulas reproduce Hunt’s classic expressions for maximum utilization and throughput. The three-station case is already far more intricate: the transition graph has eight vertices and fourteen edges, and the team used the enumeration algorithm of Gabow and Myers to detect all seventy-eight anti-arborescences that form the corresponding monomials. From these spanning polynomials, the researchers derived exact closed-form expressions not only for throughput and maximum utilization but also for work in process, the blocking probability of the second station, and the probability of the third station being idle, extending Hunt’s work to performance measures that had never before been expressed symbolically.

The four-station case demonstrates the true scale of the challenge. The transition graph contains 4,502,592 anti-arborescences, and the exact throughput formula holds 2,317,824 monomials in the numerator and 4,502,592 in the denominator, with exponents summing to a constant twenty. The researchers verified their symbolic solutions against the established MARKOV numerical algorithm across 194,481 line configurations with mean service rates between 0.9 and 1.1, finding a maximum difference of 5.37 times ten to the minus seventh, confirming that both approaches agree to within rounding error. Remarkably, evaluating the exact formula for all configurations took just over nineteen minutes on a standard desktop processor, while the MARKOV algorithm required more than eight days on the same machine.

Beyond four stations, the combinatorial explosion becomes staggering. A five-station line has a transition graph with fifty-five vertices, and the total number of anti-arborescences across all states reaches roughly 4.5 times ten to the twenty-first power. The researchers estimate that enumerating all anti-arborescences of a single state would take over a million years at their observed enumeration rate, and the whole graph would require about 140 million years. By eight stages, the count of anti-arborescences climbs to a number with over 560,000 digits, a scale that exhausts standard computer integers and demands special arbitrary-precision libraries.

The structure of the spanning polynomials also sheds new light on the well-known bowl phenomenon, the empirical observation that in balanced production lines the middle stations contribute less to overall performance than the outer ones. By recording how often each mean service rate appears as an edge weight in the transition diagrams and across all spanning trees, the team found a symmetric pattern in which the service rates of middle stages are systematically underrepresented in the formula structure. This suggests that not all stations participate equally in the exact expressions, providing a mathematical explanation for a phenomenon that has been studied since Hillier and Boling’s work in the 1960s.

Understanding the deep structure of the exact formulas also enabled the researchers to build a practical shortcut. Using genetic programming, an evolutionary artificial intelligence technique for symbolic regression, they trained a model on exact throughput values for balanced lines up to twelve stages and verified it against published values for thirteen stages. The resulting simple formula estimates throughput for perfectly balanced lines up to thirteen stages accurately to the sixth decimal digit, demonstrating that insights into formula structure can guide AI toward compact, high-accuracy approximations.

The enormous size of the exact formulas redefines what closed-form means in practice, and the authors argue that this scale is both a problem and an opportunity. On one hand, it highlights the need for computational support and special libraries for lines longer than four stages without buffers. On the other, it opens the door to innovative mathematical approaches, benchmarking problems for emerging computing systems including quantum computers, and a deeper partnership between human intuition and artificial intelligence in deriving precise mathematical relationships with reduced complexity. The work stands as a vivid reminder that even the simplest factory lines hide staggering mathematical depth, and that cracking that depth requires patience, graph theory, and a willingness to let computers enumerate millions of spanning trees in search of truth.

Subject of Research: Mathematics

Article Title: Extraction of exact symbolic stationary probability formulas for Markov chains in finite space with application to production lines. part II: unveiling accurate formulas for very short serial production lines without buffers (three- and four-stations)

Article References: Boulas, K. S., Dounias, G. D., & Papadopoulos, C. T. (2025). Extraction of exact symbolic stationary probability formulas for Markov chains in finite space with application to production lines. part II: unveiling accurate formulas for very short serial production lines without buffers (three- and four-stations). Frontiers in Manufacturing Technology, 5, Article 1439429. https://doi.org/10.3389/fmtec.2025.1439429

Image Credits: AI Generated

DOI: 10.3389/fmtec.2025.1439429

Keywords: Markov chains, production lines, throughput, spanning trees, bowl phenomenon, genetic programming, manufacturing systems, state explosion, Extraction, exact, symbolic, stationary