An integrated development environment for probabilistic relational reasoning

Abstract

This paper presents KReator, a versatile integrated development environment for probabilistic inductive logic programming currently under development. The area of probabilistic inductive logic programming (or statistical relational learning) aims at applying probabilistic methods of inference and learning in relational or first-order representations of knowledge. In the past ten years the community brought forth a lot of proposals to deal with problems in that area which mostly extend existing propositional probabilistic methods like Bayes Nets and Markov Networks on relational settings. Only few developers provide prototypical implementations of their approaches and the existing applications are often difficult to install and to use. Furthermore, due to different languages and frameworks used for the development of different systems the task of comparing various approaches becomes hard and tedious. KReator aims at providing a common and simple interface for representing, reasoning, and learning with different relational probabilistic approaches. It is a general integrated development environment which enables the integration of various frameworks within the area of probabilistic inductive logic programming and statistical relational learning. Currently, KReator implements Bayesian logic programs, Markov logic networks, and relational maximum entropy under grounding semantics. More approaches will be implemented in the near future or can be implemented by researchers themselves as KReator is open-source and available under public license. In this paper, we provide some background on probabilistic inductive logic programming and statistical relational learning and illustrate the usage of KReator on several examples using the three approaches currently implemented in KReator. Furthermore, we give an overview on its system architecture.

On probabilistic inference in relational conditional logics

“…Consider the clauses (flies(X) | bird (X)) (with a probability of 0.9 that flies(X) is true given that bird (X) is true) and (flies(X) | penguin(X)) (with a probability of 0.01 that flies(X) is true given that penguin(X) is true), cf. [11]. Using noisy-or as combining rules for flies yields a probability of approximately 0.476 for an actual penguin-bird, a much more higher probability than intended.…”

“…For the propositional case [15,20] It remains to define a satisfaction relation for conditionals with variables (see Example 2.4). Taking a naïve approach by grounding all conditionals in R universally and taking this grounding R as a propositional knowledge base, we can (usually) not determine any probability distribution that satisfies R due to its inherent inconsistency [13,11]. In the following, we propose two different approaches for semantics of (L Σ | L Σ ) prob that coincide with (2.1) in the propositional case but differ on the interpretation of population-based statements.…”

“…By defining P in this way, worlds that violate fewer instances of formulas are more probable than worlds that violate more instances (depending on the weights of the different formulas). In contrast to the approaches discussed in the previous subsection MLNs do not suffer from conflicts that arise in grounding knowledge bases as the weights of an MLN have no probabilistic interpretation, see [11] for some discussion. But from the view of knowledge representation this property is unintuitive and also distinguishes MLNs from our approach where values of conditionals do have a probabilistic interpretation.…”

“…So, grounding has to take further constraints into account, to return a consistent knowledge base (cf. [11]). While most of the above mentioned formalisms for statistical relational learning or inductive logic programming like Markov Logic Networks and Bayesian Logic Programs extend traditional (propositional) graphical models for probabilistic knowledge representation like Markov Nets and Bayesian Networks, here, we employ probabilistic conditional logic [20,32,31,26].…”

Relational Probabilistic Conditionals and Their Instantiations under Maximum Entropy Semantics for First-Order Knowledge Bases

“…In [20], a set of transformation rules PU is presented allowing to transform any consistent knowledge base R into a parametrically uniform knowledge base PU(R) with the same maximum entropy model under grounding semantics. An implementation of PU [21] is available within the KREATOR environment (KREATOR can be found at http://kreator-ide.sourceforge.net/), an integrated development environment for relational probabilistic logic [22]. The CSPU (Conditional Structures and Parametric Uniformity) component [23] of KREATOR generates PU transformation protocols, and a part the protocol for the misanthrope knowledge base R MI from Example 3 is shown in Figure 1.…”