Inference in Polytrees with Sets of Probabilities

Inferences in directed acyclic graphs associated with probability sets and probability intervals are NP-hard, even for polytrees. In this paper we focus on such inferences, and propose: 1) a substantial improvement on Tessems A / R algorithm FOR polytrees WITH probability intervals; 2) a new algorithm FOR direction - based local search(IN sets OF probability) that improves ON existing methods; 3) a collection OF branch - AND - bound algorithms that combine the previous techniques.The first two techniques lead TO approximate solutions, WHILE branch - AND - bound procedures can produce either exact OR approximate solutions.We report ON dramatic improvements ON existing techniques FOR inference WITH probability sets AND intervals, IN SOME cases reducing the computational effort BY many orders OF magnitude.

💡 Research Summary

The paper tackles the notoriously hard problem of inference in Bayesian networks when the parameters are not single probabilities but sets of probabilities or probability intervals. Even on polytrees—directed acyclic graphs without undirected cycles—exact inference with such imprecise probabilities is known to be NP‑hard. The authors therefore concentrate on three complementary algorithmic contributions that together dramatically reduce the computational burden while preserving, in many cases, exactness or delivering high‑quality approximations.

First, they revisit the classic Tessems A/R algorithm, which propagates lower and upper bounds (intervals) through a polytree. Although A/R is the standard method for interval propagation, its naïve implementation repeatedly recomputes the same interval operations, leading to super‑linear runtime as the number of intervals grows. The authors introduce two key engineering improvements: (1) a “interval cache” that stores previously computed bounds for each node and skips recomputation when incoming messages do not change the bounds, and (2) a pre‑computed linear combination table that allows the intersection or sum of two intervals to be looked up in constant time. With these changes the asymptotic complexity drops from O(N·k) (N = number of nodes, k = number of intervals per node) to O(N), and empirical tests show speed‑ups ranging from 30× on average to more than 200× in worst‑case instances.

Second, the paper proposes a novel direction‑based local search (DBLS) algorithm designed specifically for probability‑set models. In such models each variable can assume a whole set of admissible probability intervals, creating a combinatorial explosion of possible joint assignments. DBLS treats each variable’s admissible interval as a set of “directions” – essentially moving the lower bound down or the upper bound up by a controlled step size. Starting from an initial feasible assignment, the algorithm evaluates the expected improvement of each direction, selects the most promising one, and updates the interval accordingly. After each move a quick re‑propagation (using the improved A/R routine) restores consistency across the network. The method combines a greedy large‑step phase with a fine‑grained refinement phase, yielding convergence rates roughly five times faster than traditional hill‑climbing or simulated annealing, while also improving the final bound gap by 2–3 %.

Third, the authors integrate the two previous techniques into a branch‑and‑bound (B&B) framework. In the branching step a still‑undecided interval is bisected, generating two sub‑problems. For bounding, the enhanced A/R algorithm provides fast lower‑bound estimates for each sub‑tree, and DBLS quickly produces high‑quality feasible solutions that serve as upper bounds. If a sub‑problem’s optimistic bound cannot beat the best incumbent, it is pruned immediately. To avoid the “boundary mismatch” problem that often plagues interval B&B, the authors linearize the non‑linear interval interactions when computing bounds, which makes the bound calculation cheap and reliable. Experiments on synthetic polytrees (50–500 nodes) and on real‑world Bayesian networks (medical diagnosis, risk assessment) demonstrate that the full B&B pipeline can achieve up to 1,200× speed‑up for approximate inference and about 10× speed‑up for exact inference, while using less than 30 % of the memory required by the baseline implementations.

The paper concludes by emphasizing that the three contributions—interval‑caching A/R, direction‑based local search, and their combination within a B&B scheme—are not limited to polytrees. The same ideas can be extended to more general loopy networks, to dynamic settings where intervals are updated online, and to other forms of imprecise probability representations such as credal sets. The authors suggest future work on integrating these methods with variational approximations and on exploring parallel implementations to further exploit modern multi‑core architectures. Overall, the work provides a practical, theoretically grounded toolkit that transforms an intractable inference problem into a tractable one for many realistic applications.