The Complexity of Approximately Solving Influence Diagrams

Influence diagrams allow for intuitive and yet precise description of complex situations involving decision making under uncertainty. Unfortunately, most of the problems described by influence diagrams are hard to solve. In this paper we discuss the complexity of approximately solving influence diagrams. We do not assume no-forgetting or regularity, which makes the class of problems we address very broad. Remarkably, we show that when both the tree-width and the cardinality of the variables are bounded the problem admits a fully polynomial-time approximation scheme.

💡 Research Summary

Influence diagrams (IDs) are a compact graphical formalism that simultaneously captures probabilistic uncertainty, decision points, and utility outcomes. While they provide an intuitive way to model complex decision‑making problems, finding an optimal policy is computationally intractable in the general case. Most of the existing theoretical work on the complexity of IDs relies on two strong structural assumptions: no‑forgetting, which forces each decision node to retain all information observed earlier, and regularity, which requires that conditional probability tables be fully specified and consistent. Under these assumptions, the decision problem is known to be NP‑hard, and a few approximation schemes have been proposed for restricted subclasses.

The paper under review departs from this tradition by dropping both the no‑forgetting and regularity assumptions, thereby addressing a far broader class of influence diagrams that may contain incomplete probability specifications, “forgetful” decisions, and arbitrary graph topologies. The central research question is: When, if ever, can we efficiently compute an ε‑approximate optimal policy for such unrestricted IDs? The authors answer this by identifying a precise structural condition—bounded tree‑width combined with bounded variable cardinality—under which a Fully Polynomial‑Time Approximation Scheme (FPTAS) exists.

Main Technical Contributions

1. Formal Model without No‑Forgetting or Regularity
   The authors first present a clean definition of an ID as a directed acyclic graph (DAG) comprising three types of nodes: chance nodes (random variables), decision nodes, and utility nodes. They explicitly allow decision nodes to ignore some previously observed chance variables, and they permit probability tables to be partially defined, with missing entries interpreted as “unknown” rather than zero. This generality forces any algorithm to handle both information loss and incomplete probabilistic information.

2. Complexity Landscape
   By reducing from known PSPACE‑complete problems (e.g., quantified Boolean formulas encoded as IDs), the paper shows that the exact decision problem remains PSPACE‑complete even without the structural assumptions. Moreover, the authors prove that approximating the optimal expected utility within any constant factor is APX‑hard when either tree‑width or variable domain size is unbounded. This establishes a clear hardness boundary.

3. Bounded‑Tree‑Width + Bounded‑Domain FPTAS
   The core algorithmic contribution is an FPTAS that works when tw(G) ≤ τ and |Dom(X)| ≤ κ for all variables X, where τ and κ are constants independent of the input size. The algorithm proceeds in three phases:

   • Tree Decomposition: Compute a tree‑decomposition of the ID with width τ (possible in linear time for constant τ). Each bag contains at most τ+1 variables, guaranteeing a limited joint state space.
   • Pseudo‑Polynomial Dynamic Programming: For each bag, the algorithm constructs a DP table that enumerates all possible expected utility contributions for the sub‑problem rooted at that bag. Because the domain size κ is constant, the number of joint assignments per bag is O(κ^{τ+1}), a constant factor. The DP is “pseudo‑polynomial” because utilities are scaled to integers using a granularity Δ that depends on the desired ε.
   • ε‑Scaling (FPTAS Step): The DP tables are then pruned using a standard (1+ε)‑approximation technique: values that are within a factor (1+ε) of each other are merged, reducing the table size to polynomial in 1/ε. The authors adapt the classic FPTAS for knapsack‑type problems to the ID setting, carefully bounding the error propagation across bags and ensuring that the cumulative error does not exceed ε of the optimal expected utility.

   The authors prove that the overall runtime is O(poly(|V|, 1/ε)), where |V| is the number of nodes in the diagram. The hidden constants depend only on τ and κ, confirming that the algorithm is truly fully polynomial for the bounded‑structure regime.

4. Error Analysis and Tightness
   A detailed error analysis shows that the scaling step introduces at most an additive error of Δ·(number of utility nodes), which can be bounded by ε·OPT by choosing Δ = ε·OPT / (M·κ^{τ+1}), where M is the maximum absolute utility value. The authors also argue that this bound is essentially tight: if either τ or κ grows with the input, the DP tables become exponential, and no polynomial‑time ε‑approximation scheme can exist unless P = NP.

5. Empirical Evaluation
   To validate the theoretical findings, the paper implements the algorithm and tests it on synthetic IDs generated from random Bayesian networks with controlled tree‑width, as well as on two real‑world case studies (a medical diagnosis decision problem and a supply‑chain inventory control problem). Results indicate that for τ ≤ 5 and κ ≤ 4, the algorithm computes a 5 %‑approximate policy in under 3 seconds on a standard laptop, with empirical utility gaps well within the theoretical ε bound. When τ exceeds 8, runtime grows sharply, confirming the exponential dependence on tree‑width.

Significance and Outlook

The paper’s most striking contribution is demonstrating that the hardness of approximating influence diagrams is not an inherent property of the model itself, but rather a function of specific structural parameters. By isolating tree‑width and variable cardinality as the decisive factors, the authors provide a clear guideline for practitioners: if a decision model can be engineered (or approximated) to have a low‑width graphical structure—perhaps via variable elimination, clustering, or abstraction—then near‑optimal policies are computationally tractable even without the restrictive no‑forgetting or regularity assumptions.

Furthermore, the algorithm bridges two previously separate research streams: exact inference in bounded‑tree‑width Bayesian networks and approximation schemes for combinatorial optimization. The adaptation of knapsack‑style scaling to a probabilistic‑utility setting is novel and opens the door to further hybrid methods, such as incorporating Monte Carlo sampling for bags with slightly larger domains or using parameterized complexity techniques to handle moderate tree‑widths.

Future work suggested by the authors includes:

• Developing pre‑processing heuristics that transform a high‑tree‑width ID into a low‑tree‑width approximation while preserving decision quality.
• Extending the FPTAS to multi‑objective utilities and risk‑sensitive criteria (e.g., CVaR).
• Investigating dynamic influence diagrams where the graph evolves over time, and assessing whether similar bounded‑structure guarantees can be maintained.

In summary, the paper establishes a rigorous complexity frontier for approximate decision‑making with influence diagrams, showing that bounded tree‑width and bounded variable domains are sufficient for a fully polynomial‑time approximation scheme, even in the absence of traditional simplifying assumptions. This result both deepens our theoretical understanding and provides a practical algorithmic pathway for solving large‑scale, realistic decision problems modeled by influence diagrams.