Credal Valuation Networks

• Credal valuation networks are graphical models that use credal sets—closed convex sets of probability distributions—to generalize Bayesian networks for imprecise inference.
• They integrate operations like combination and marginalization via convex optimization and polyhedral geometry, ensuring sound and exact marginal interval computations.
• Their framework connects valuation-based systems with robust inference methods, offering scalable algorithms while addressing NP-hard challenges in uncertainty reasoning.

A credal valuation network is a graphical model for reasoning under uncertainty, in which the local uncertainties are represented by credal sets—closed convex sets of probability distributions—rather than precise probabilities. This framework generalizes valuation-based systems and Bayesian networks by allowing coherent imprecision at the level of local potentials and their combinations. Operations such as combination and marginalization are defined so as to extend classical probabilistic inference to credal (imprecise-probability) settings, while satisfying the axioms of valuation algebras. Credal valuation networks support exact, sound inference of marginal intervals in the presence of epistemic or systemic uncertainty, and their analysis connects convex optimization, polyhedral geometry, and algorithmic advances in graphical models.

1. Mathematical Foundations: Credal Sets and Valuation Algebra

A credal set over a finite variable $X$ is a closed convex subset of probability mass functions on the domain $\Theta_X$:

$$K^X = \{\,p \in \mathcal{P}^X : \underline p_i \le p(x_i) \le \overline p_i,\,\forall i \;|\; \sum_{i=1}^n p(x_i) = 1 \},$$

where $\underline p_i,\overline p_i$ are lower and upper bounds that satisfy coherence constraints ensuring the set is nonempty and convex. This interval-based specification generalizes the case $\underline p_i = \overline p_i$ (precise PMF).

A valuation network comprises variable-nodes and valuation-nodes, the latter associated with credal sets over subsets of variables. The set $\{\Phi, d, \otimes, \downarrow\}$ (valuations, domains, combination, marginalization) forms a valuation algebra when:

• Combination ($\otimes$): Given $K_1^{D}$, $K_2^{D}$ (credal sets on domain $D$),

$\Theta_X$0

where bounds are determined, for each $\Theta_X$1, by extremal normalized products over pairs of distributions in $\Theta_X$2, $\Theta_X$3.

• Marginalization ($\Theta_X$4): For $\Theta_X$5 and $\Theta_X$6, the marginal is

$\Theta_X$7

with bounds computed via interval arithmetic and tight reduction.

The combination/marginalization operators, along with vacuous extension to joint domains, satisfy the standard axioms (closure, associativity, commutativity, distributivity, identity), confirming that credal valuation networks are consistent valuation algebras (Ristic et al., 2022).

2. Structural and Algorithmic Aspects of Credal Valuation Networks

A credal valuation network is constructed as an undirected bipartite graph. Variable nodes $\Theta_X$8 are linked through valuation nodes $\Theta_X$9, each attached to a credal set over a domain $$K^X = \{\,p \in \mathcal{P}^X : \underline p_i \le p(x_i) \le \overline p_i,\,\forall i \;|\; \sum_{i=1}^n p(x_i) = 1 \},$$0. Edges connect $$K^X = \{\,p \in \mathcal{P}^X : \underline p_i \le p(x_i) \le \overline p_i,\,\forall i \;|\; \sum_{i=1}^n p(x_i) = 1 \},$$1 to all $$K^X = \{\,p \in \mathcal{P}^X : \underline p_i \le p(x_i) \le \overline p_i,\,\forall i \;|\; \sum_{i=1}^n p(x_i) = 1 \},$$2 in $$K^X = \{\,p \in \mathcal{P}^X : \underline p_i \le p(x_i) \le \overline p_i,\,\forall i \;|\; \sum_{i=1}^n p(x_i) = 1 \},$$3. The network globally defines a convex polytope of joint PMFs consistent with all local credal constraints.

Inference proceeds by a local computation or fusion algorithm: variables not in the marginal query are sequentially eliminated by fusing all credal sets involving the variable, combining them into a joint credal set over the full relevant scope, then marginalizing out the eliminated variable. The result is a sequence of local operations, ensuring exact marginal interval inference for any subset of variables.

Implementation exploits properties such as:

• Storing valuations as pairs of vectors $$K^X = \{\,p \in \mathcal{P}^X : \underline p_i \le p(x_i) \le \overline p_i,\,\forall i \;|\; \sum_{i=1}^n p(x_i) = 1 \},$$4
• Performing marginalizations and vacuous extensions efficiently via linear algebra
• Reducing combination to a series of min/max bilinear programs (or QPs)
• Exploiting fast linear or convex programming for precise/mixed cases

The architecture permits further efficiency via binary join trees and message passing. These aspects enable scalability to moderately sized domains, with each credal-fusion step limited by the size of the largest local domain (Ristic et al., 2022).

3. Polyhedral Geometry, Composition, and Network Construction

In the general valuation-based systems (VBS) paradigm, multidimensional credal models are built from low-dimensional credal sets via a composition operator and marginalization (Vejnarová et al., 2017). Each local credal set $$K^X = \{\,p \in \mathcal{P}^X : \underline p_i \le p(x_i) \le \overline p_i,\,\forall i \;|\; \sum_{i=1}^n p(x_i) = 1 \},$$5 is represented both in $$K^X = \{\,p \in \mathcal{P}^X : \underline p_i \le p(x_i) \le \overline p_i,\,\forall i \;|\; \sum_{i=1}^n p(x_i) = 1 \},$$6- and $$K^X = \{\,p \in \mathcal{P}^X : \underline p_i \le p(x_i) \le \overline p_i,\,\forall i \;|\; \sum_{i=1}^n p(x_i) = 1 \},$$7-representation (vertices and inequalities).

Given two credal sets $$K^X = \{\,p \in \mathcal{P}^X : \underline p_i \le p(x_i) \le \overline p_i,\,\forall i \;|\; \sum_{i=1}^n p(x_i) = 1 \},$$8 and $$K^X = \{\,p \in \mathcal{P}^X : \underline p_i \le p(x_i) \le \overline p_i,\,\forall i \;|\; \sum_{i=1}^n p(x_i) = 1 \},$$9, their polyhedral composition $\underline p_i,\overline p_i$0 on $\underline p_i,\overline p_i$1 is defined as the convex hull of joint PMFs $\underline p_i,\overline p_i$2 such that the marginals of $\underline p_i,\overline p_i$3 on $\underline p_i,\overline p_i$4 and $\underline p_i,\overline p_i$5 belong to $\underline p_i,\overline p_i$6 and $\underline p_i,\overline p_i$7, respectively, and such that, on the intersection $\underline p_i,\overline p_i$8, the marginals match as closely as possible. When marginals cannot be matched exactly, vacuous extension is used for the non-projective part.

Algorithmically, composition reduces to projecting vertices, solving QPs to find minimal-distance marginals, and maintaining all constructed extreme points, with convex hull operations to unify the result. Marginalization is implemented via linear projection of the polytope. The global credal set is constructed by an ordered sequence of compositions, optionally followed by normalization and projection for answering marginal queries (Vejnarová et al., 2017).

4. Inference Complexity and Algorithmic Hardness

Exact inference in credal networks, even for polytrees under separate specification, is NP-hard. The canonical decision problem—determining if the lower probability $\underline p_i,\overline p_i$9 exceeds a threshold—can be shown NP-complete via reduction from the SubsetSum problem (Rocha et al., 2012).

Despite this, for special cases such as binary-valued polytrees, context-wise separability can reduce the complexity to polynomial time (recovering the 2U-algorithm). Separable variable elimination algorithms exploit context decomposition, terminal evidence pruning, bucket elimination with context-wise convex hull reduction, and careful ordering to manage computational burdens while maintaining exactness.

For large or highly connected networks, the number of extreme points—and consequently the size of intermediate credal sets—may become exponential in the size of separator domains, limiting practical scalability. Convex hull and projection computations represent the key bottlenecks, motivating algorithmic approximations or relaxations in more challenging scenarios (Rocha et al., 2012, Vejnarová et al., 2017).

5. Connections to Credal Bayesian Networks and Robustness

Credal valuation networks subsume credal Bayesian networks (CrBNs), which specify interval-based credal sets for each conditional probability table row. Inference in CrBNs can be reformulated as constrained polynomial optimization over arithmetic or sum–product circuits. By relaxing equality constraints across duplicated credal sets in the network polynomial's circuit representation, linear-time upper bounds on maximum marginal probabilities (MAR$\underline p_i = \overline p_i$0) can be efficiently computed. These relaxations yield robustness guarantees by bounding the worst-case event probabilities and have provable tightness analyses in terms of duplicity of credal set instances in the compiled circuits (Wijk et al., 2022).

This framework delivers fast, sound, and reasonably tight bounds in practice, with the constraint relaxation guaranteeing upper bounds that are exact in tree-like or singly-occuring credal set settings.

6. Numerical Illustration and Implementation Practices

A demonstrative example is the “Arrival Delay” network with seven variables and credal sets given as tight intervals around the empirical PMFs. Marginalization of the arrival delay variable produces intervals that tightly sandwich the ground-truth PMF, while comparison to Dempster–Shafer evidential networks reveals that credal valuation networks often yield tighter probability intervals (Ristic et al., 2022).

Practically, implementation requires:

• Efficient storage and manipulation of $\underline p_i = \overline p_i$1 vectors
• Use of convex or quadratic programming for combination steps involving general credal sets
• Special-case optimizations if one credal set is precise (reducing the calculation to LP)
• Lazy computation, warm-start strategies, and ordering to exploit problem structure

For small-scale or sparsely connected problems, inference is tractable. For high-dimensional domains, polyhedral computations may be prohibitively expensive, motivating the exploration of approximation algorithms or structural simplifications.

7. Theoretical and Practical Significance

Credal valuation networks provide a rigorous framework for reasoning under bounded rationality, epistemic uncertainty, and imprecise knowledge, subsuming both precise probabilistic inference and broader imprecise-probability methodologies. Their theoretical foundations in convex geometry, polyhedral operations, and valuation algebra guarantee soundness and extendability to a variety of application contexts.

Open theoretical questions focus on optimal distance selection in polyhedral composition, efficient management of exponential growth in extremal points, and the design of algorithms that trade off exactness and tractability. Practically, their robust, interpretable interval outputs provide principled bounds for decision support in adversarial, data-poor, or high-stakes settings.

References:

• (Rocha et al., 2012): Inference with Separately Specified Sets of Probabilities in Credal Networks
• (Ristic et al., 2022): Credal Valuation Networks for Machine Reasoning Under Uncertainty
• (Vejnarová et al., 2017): Composition of Credal Sets via Polyhedral Geometry
• (Wijk et al., 2022): Robustness Guarantees for Credal Bayesian Networks via Constraint Relaxation over Probabilistic Circuits