Bayesian networks and causal models provide frameworks for reasoning about external interventions, enabling tasks that go beyond what probability distributions alone can support. Although these formalisms are often informally described as encoding causal knowledge, there is a lack of a formal theory that characterizes the kind of knowledge required to predict the effects of such interventions. This work introduces the theoretical framework of causal systems to implement Aristotle’s distinction between knowledge- that and knowledge- why within the setting of artificial intelligence. By interpreting existing AI technologies as causal systems, it examines the corresponding forms of knowledge they embody. Finally, it argues that predicting the effects of external interventions is possible only with knowledge- why , offering a more precise account of the assumptions underlying this capacity.

Counterfactual reasoning describes the ability to consider alternative outcomes – to answer questions such as “What if I had been nicer? Would more students have attended my lecture?” Counterfactuals arise in a wide range of applications; for instance, ethical notions such as responsibility, blame, and harm can be defined in counterfactual terms. Pearl’s standard approach to evaluating counterfactuals requires a fully specified causal model. Such models are difficult to specify, and counterfactual reasoning remains underdetermined when only a probability distribution or a Bayesian network is given. This work proposes a novel approach to counterfactual reasoning in Boolean Bayesian networks. Drawing inspiration from Lewis’s account of counterfactuals and the principle of maximum entropy, it argues that a rational agent should assess counterfactuals by explaining the network through a causal model derived from solving an optimization problem. This problem is formulated in terms of distances between actual and alternative realities, and each counterfactual corresponds to one such problem. The main result identifies a unique causal model that simultaneously solves all of these optimization problems. This provides a principled justification for counterfactual reasoning based solely on a Boolean Bayesian network, without requiring the full specification of a causal model.

Abstract Extrapolation with domain size has received plenty of attention recently, both in its own right and as part of the broader issue of scaling inference and learning to large domains. We consider Markov logic networks and relational logistic regression as two fundamental representation formalisms in statistical relational artificial intelligence that use weighted formulas in their specification. However, Markov logic networks are based on undirected graphs, while relational logistic regression is based on directed acyclic graphs. We show that when scaling the weight parameters with the domain size, the asymptotic behaviour of a relational logistic regression model is transparently controlled by the parameters, and we supply an algorithm to compute asymptotic probabilities. We show using two examples that this is not true for Markov logic networks. We also discuss using several examples, mainly from the literature, how the application context can help the user to decide when such scaling is appropriate and when using the raw unscaled parameters might be preferable.

Dependencies on the relative frequency of a state in the domain are common when modelling probabilistic dependencies on relational data. For instance, the likelihood of a school closure during an epidemic might depend on the proportion of infected pupils exceeding a threshold. Often, rather than depending on discrete thresholds, dependencies are continuous: for instance, the likelihood of any one mosquito bite transmitting an illness depends on the proportion of carrier mosquitoes. Current approaches usually only consider probabilities over possible worlds rather than over domain elements themselves. An exception are the recently introduced Lifted Bayesian Networks for Conditional Probability Logic, which express discrete dependencies on probabilistic data. We introduce functional lifted Bayesian networks, a formalism that explicitly incorporates continuous dependencies on relative frequencies into statistical relational artificial intelligence. and compare and contrast them with ifted Bayesian Networks for Conditional Probability Logic. Incorporating relative frequencies is not only beneficial to modelling; it also provides a more rigorous approach to learning problems where training and test or application domains have different sizes. To this end, we provide a representation of the asymptotic probability distributions induced by functional lifted Bayesian networks on domains of increasing sizes. Since that representation has well-understood scaling behaviour across domain sizes, it can be used to estimate parameters for a large domain consistently from randomly sampled subpopulations. Furthermore, we show that in parametric families of FLBN, convergence is uniform in the parameters, which ensures a meaningful dependence of the asymptotic probabilities on the parameters of the model.

We generalise the distribution semantics underpinning probabilistic logic programming by distilling its essential concept, the separation of a free random component and a deterministic part. This abstracts the core ideas beyond logic programming as such to encompass frameworks from probabilistic databases, probabilistic finite model theory and discrete lifted Bayesian networks. To demonstrate the usefulness of such a general approach, we completely characterise the projective families of distributions representable in the generalised distribution semantics and we demonstrate both that large classes of interesting projective families cannot be represented in a generalised distribution semantics and that already a very limited fragment of logic programming (acyclic determinate logic programs) in the deterministic part suffices to represent all those projective families that are representable in the generalised distribution semantics at all.

We consider logics with truth values in the unit interval $[0,1]$. Such logics are used to define queries and to define probability distributions. In this context the notion of almost sure equivalence of formulas is generalized to the notion of asymptotic equivalence. We prove two new results about the asymptotic equivalence of formulas where each result has a convergence law as a corollary. These results as well as several older results can be formulated as results about the relative asymptotic expressivity of inference frameworks. An inference framework $\mathbf{F}$ is a class of pairs $(\mathbb{P}, L)$, where $\mathbb{P} = (\mathbb{P}_n : n = 1, 2, 3, \ldots)$, $\mathbb{P}_n$ are probability distributions on the set $\mathbf{W}_n$ of all $\sigma$-structures with domain $\{1, \ldots, n\}$ (where $\sigma$ is a first-order signature) and $L$ is a logic with truth values in the unit interval $[0, 1]$. An inference framework $\mathbf{F}'$ is asymptotically at least as expressive as an inference framework $\mathbf{F}$ if for every $(\mathbb{P}, L) \in \mathbf{F}$ there is $(\mathbb{P}', L') \in \mathbf{F}'$ such that $\mathbb{P}$ is asymptotically total variation equivalent to $\mathbb{P}'$ and for every $\varphi(\bar{x}) \in L$ there is $\varphi'(\bar{x}) \in L'$ such that $\varphi'(\bar{x})$ is asymptotically equivalent to $\varphi(\bar{x})$ with respect to $\mathbb{P}$. This relation is a preorder. If, in addition, $\mathbf{F}$ is at least as expressive as $\mathbf{F}'$ then we say that $\mathbf{F}$ and $\mathbf{F}'$ are asymptotically equally expressive. Our third contribution is to systematize the new results of this paper and several previous results in order to get a preorder on a number of inference systems that are of relevance in the context of machine learning and artificial intelligence.

For a finite and relational signature σ and finite domain D we consider the set WD of all σ-structures with domain D. On WD a probability distribution is determined by a so-called parametrized probabilistic graphical model, a concept studied in statistical relational artificial intelligence. We also consider a many valued logic, denoted PLA, with truth values in the unit interval for expressing queries. PLA uses aggregation functions, for example the arithmetic mean, geometric mean, maximum and minimum, instead of quantifiers. In this setting we prove that every formula of PLA with only admissible aggregation functions is asymptotically equivalent to a formula without aggregation functions, as the domain size tends to infinity. A corollary of this is a probabilistic convergence law for PLA-formulas with only admissible aggregation functions.

Abstract A ProbLog program is a logic program with facts that only hold with a specified probability. In this contribution, we extend this ProbLog language by the ability to answer “What if” queries. Intuitively, a ProbLog program defines a distribution by solving a system of equations in terms of mutually independent predefined Boolean random variables. In the theory of causality, Judea Pearl proposes a counterfactual reasoning for such systems of equations. Based on Pearl’s calculus, we provide a procedure for processing these counterfactual queries on ProbLog programs, together with a proof of correctness and a full implementation. Using the latter, we provide insights into the influence of different parameters on the scalability of inference. Finally, we also show that our approach is consistent with CP-logic, that is with the causal semantics for logic programs with annotated with disjunctions.

Abstract Probabilistic logic programming is a major part of statistical relational artificial intelligence, where approaches from logic and probability are brought together to reason about and learn from relational domains in a setting of uncertainty. However, the behaviour of statistical relational representations across variable domain sizes is complex, and scaling inference and learning to large domains remains a significant challenge. In recent years, connections have emerged between domain size dependence, lifted inference and learning from sampled subpopulations. The asymptotic behaviour of statistical relational representations has come under scrutiny, and projectivity was investigated as the strongest form of domain size dependence, in which query marginals are completely independent of the domain size. In this contribution we show that every probabilistic logic program under the distribution semantics is asymptotically equivalent to an acyclic probabilistic logic program consisting only of determinate clauses over probabilistic facts. We conclude that every probabilistic logic program inducing a projective family of distributions is in fact everywhere equivalent to a program from this fragment, and we investigate the consequences for the projective families of distributions expressible by probabilistic logic programs.