Evaluating influence diagrams with decision circuits

Although a number of related algorithms have been developed to evaluate influence diagrams, exploiting the conditional independence in the diagram, the exact solution has remained intractable for many important problems. In this paper we introduce decision circuits as a means to exploit the local structure usually found in decision problems and to improve the performance of influence diagram analysis. This work builds on the probabilistic inference algorithms using arithmetic circuits to represent Bayesian belief networks [Darwiche, 2003]. Once compiled, these arithmetic circuits efficiently evaluate probabilistic queries on the belief network, and methods have been developed to exploit both the global and local structure of the network. We show that decision circuits can be constructed in a similar fashion and promise similar benefits.

💡 Research Summary

The paper tackles the long‑standing computational bottleneck in exact evaluation of Influence Diagrams (IDs) by introducing Decision Circuits, a data structure that extends the well‑known Arithmetic Circuits (ACs) used for Bayesian network inference. An AC is a directed acyclic graph whose internal nodes perform multiplication (for factor combination) and addition (for marginalisation). By compiling a Bayesian network into an AC, one can answer probabilistic queries in time proportional to the size of the circuit, which is often dramatically smaller than the full joint distribution because the circuit exploits global conditional independence and local factorisation.

Decision Circuits augment this framework with a third type of internal node – a max node – to represent decision variables. In an ID, each decision node selects an action that maximises expected utility given the information available at that point. By embedding max nodes directly into the circuit, the algorithm can compute both the optimal policy and its expected utility in a single pass, avoiding the separate dynamic‑programming recursions traditionally required. Utility nodes are treated as leaf factors that feed into the circuit, allowing the expected utility to be accumulated alongside probability calculations.

The compilation process proceeds in three stages. First, the ID is transformed into a normal form: decision nodes are ordered to reflect the information precedence, and the utility function is expressed as a linear combination of elementary utilities. Second, a variable‑elimination order is selected. Unlike pure probabilistic inference, the order must balance two competing goals: minimise the tree‑width of the probabilistic sub‑graph and keep the sub‑graph containing max nodes as sparse as possible. The authors adapt existing tree‑width minimisation heuristics and introduce a locality‑aware scoring function that penalises orders that would create large max‑clusters. Third, the circuit is built by traversing the chosen order: multiplication nodes combine conditional probability tables, addition nodes marginalise out chance variables, and max nodes perform a local optimisation over the decision’s action space. Because each sub‑circuit is reused wherever the same factor appears, the final structure is often orders of magnitude smaller than the naïve decision‑tree representation.

Evaluation of a compiled Decision Circuit consists of a forward pass and a backward pass. In the forward pass, the circuit is evaluated from the leaves to the root, applying the three operators in topological order. The root value yields the maximal expected utility. The backward pass propagates partial derivatives (or sensitivities) from the root back to the decision nodes, thereby identifying the action that achieved the max at each decision point. This dual‑pass scheme simultaneously produces the optimal policy and its value, with no need for separate policy‑extraction steps.

Empirical results on a suite of benchmark IDs – ranging from simple medical diagnosis models to larger supply‑chain planning problems – demonstrate that Decision Circuits achieve substantial reductions in both time and memory compared to traditional exact algorithms such as variable‑elimination with policy trees or the Shenoy‑Shafer architecture. In many cases the compiled circuit size grows only linearly with the number of decisions, whereas the policy‑tree size grows exponentially. The authors report speed‑ups of 5‑ to 30‑fold and memory savings of up to 80 % while preserving exactness of the solution.

The paper’s contributions can be summarised as follows:

1. Formal definition of Decision Circuits, integrating max operators into the arithmetic‑circuit paradigm.
2. Compilation algorithm that jointly exploits global conditional independence and local decision‑structure, yielding compact circuits.
3. Demonstration of exact, efficient evaluation of IDs, with empirical evidence of superior scalability over existing exact methods.

Future work outlined includes incremental recompilation for dynamic evidence updates, handling of non‑linear or multi‑objective utilities, and distributed/parallel execution of large circuits. These extensions would broaden the applicability of Decision Circuits to real‑time decision‑support systems, autonomous agents, and large‑scale operational planning.