Abstract
Bayesian networks provide an elegant formalism for representing and reasoning about uncertainty using probability theory. However, as a probabilistic extension of propositional logic, they inherit limitations regarding the representation of objects and complex relations. We introduce Bayesian Logic Programs (BLPs), a generalization that combines Bayesian networks with definite clause logic. By establishing a one-to-one mapping between ground atoms and random variables, BLPs maintain a strict separation between qualitative structure and quantitative probability. This framework generalizes both Bayesian networks and logic programs, offering a powerful tool for modeling relational domains with inherent uncertainty.
Introduction
Traditional Bayesian networks (BNs) are fundamentally "propositional." They excel at modeling dependencies within a fixed set of variables but struggle in dynamic or relational environments, such as computer networks with variable nodes or complex genealogical structures. In these scenarios, a standard BN must be manually reconstructed for every specific instance of the domain.
Logic programming (e.g., Prolog) was designed to handle such relational complexity and recursion, yet it lacks a native mechanism for probabilistic reasoning. Bayesian Logic Programs (BLPs) address this by unifying the two paradigms. In a BLP, the logical structure defines the dependencies between random variables, allowing for a flexible, template-based approach to probabilistic modeling.
The BLP Framework
Bayesian Clauses
The core building block of a BLP is the Bayesian clause. A clause is represented as A | A1, ..., An, signifying that the random variable $A$ is conditionally dependent on the variables $A_1$ through $A_n$. Unlike standard logic, where an atom is either true or false, every ground atom in a BLP (e.g., height(john)) is treated as a unique random variable that can take on values from a specific domain (discrete or continuous).
Separation of Concerns
One of the primary advantages of BLPs is the clean separation between qualitative and quantitative information:
- Qualitative Component: The set of Bayesian clauses defines the dependency structure of the model (the directed acyclic graph).
- Quantitative Component: Each Bayesian clause is associated with a Conditional Probability Distribution (CPD). This defines the specific probabilistic relationship between the head of the clause and its body.
Combining Rules
In relational domains, multiple logical derivations might lead to the same conclusion. For instance, a child's height is influenced by both the mother's and the father's height. BLPs utilize combining rules (such as noisy-or, max, or linear addition) to merge these distinct influences into a single, coherent conditional distribution for the target random variable.
Knowledge-Based Model Construction (KBMC)
To perform inference, BLPs do not require the construction of a global, exhaustive network. Instead, they employ Knowledge-Based Model Construction (KBMC). When a query is made, the system uses a backward-chaining procedure—similar to Prolog’s SLD-resolution—to identify only the relevant variables. This results in a "support network," which is the minimal Bayesian network necessary to answer the specific query. This approach allows BLPs to handle potentially infinite domains (due to recursion) as long as the support network for a given query remains finite.
Representational Power and Insights
Generalization of Existing Models
BLPs are a true superset of both Bayesian networks and pure logic programs. If a BLP consists only of facts (clauses without bodies), it reduces to a standard Bayesian network. Conversely, if the probabilities within the CPDs are strictly deterministic (1.0 or 0.0), the framework functions as a standard logic program. Furthermore, BLPs can naturally represent complex temporal models like Hidden Markov Models (HMMs) and Dynamic Bayesian Networks (DBNs) by using logical variables to represent time steps.
Handling Continuous Domains
By treating ground atoms as random variables rather than boolean states, BLPs can naturally incorporate continuous data. For example, a BLP can model quantitative genetics by representing the inheritance of height across generations using Gaussian distributions, a task that is cumbersome in purely propositional or purely logical frameworks.
Conclusion and Future Directions
Bayesian Logic Programs provide a robust declarative semantics for relational probabilistic modeling. By mapping ground atoms to random variables, they offer a scalable way to apply Bayesian reasoning to complex, structured data.
While the current framework establishes a strong foundation for representation and inference, several avenues for future work remain. These include the integration of "negation as failure" into the probabilistic context and the development of Inductive Logic Programming (ILP) techniques to automatically learn both the structure and the parameters of BLPs from data. Additionally, as models grow in complexity, the implementation of approximate inference techniques like Markov Chain Monte Carlo (MCMC) will be essential for maintaining computational efficiency.
Summary updated