An Inductive Principle for Learning Logical De nitions from Relationsy
نویسنده
چکیده
Induction for Horn clauses from data received a major boost from the success of FOIL 1;2;3 as a pioneer clausal learning algorithm. FOIL attempts to explain a target relation (while avoiding a negative set) by incrementally adding clauses that successively increase this explained part. The residue of the target set is obtained by postulating the target set in the antecedents (body) of the generated clauses. In contrast to this, if one considers the conventional least model semantics of Horn clauses, this semantics is realized operationally by a xed point computation from the set of facts. In this paper we consider a variant of FOIL in which the clause search is like in FOIL ? but the residue of the target is computed diierently. This residue computation is motivated by xed point theory i.e. the residue of target set is obtained by using the generated (explained) part of target set in the antecedents (body) of the generated clauses. The current generated clauses together with the initial facts or tuples are used to compute a closure, i.e., its current xed point. The target minus this current xed point is the residue. As a result of this shift, we have a monotonic sequence of xed points such that if termination is achieved the target will automatically be the xed point of the nal set of clauses, and correct cover is guaranteed.
منابع مشابه
A Type-Free Theory of Half-Monotone Inductive Definitions
This paper studies an extension of inductive de nitions in the context of a type-free theory. It is a kind of simultaneous inductive de nition of two predicates where the de ning formulas are monotone with respect to the rst predicate, but not monotone with respect to the second predicate. We call this inductive de nition half-monotone in analogy of Allen's term half-positive. We can regard thi...
متن کاملConnrmation-guided Discovery of Rst-order Rules with Tertius
This paper deals with learning rst-order logic rules from data lacking an explicit classi cation predicate. Consequently, the learned rules are not restricted to predicate de nitions as in supervised inductive logic programming. First-order logic o ers the ability to deal with structured, multi-relational knowledge. Possible applications include rst-order knowledge discovery, induction of integ...
متن کاملLinkk Oping Electronic Articles in an Inductive Deenition Approach to Ramiications Linkk Oping Electronic Articles in Computer and Information Science
In the current state of the art on the rami cation problem the purpose of causal laws is to restore the integrity of state constraints In contrast with this view we argue that causal laws should be seen as representations of how physical or logical forces and e ects propagate through a dynamic system We argue that in order to obtain a natural and modular representation of the e ect propagation ...
متن کاملUltimate Intelligence Part II: Physical Measure and Complexity of Intelligence
We investigate physical measures and limits of intelligence that are objective and useful. We propose a universal measure of operator induction fitness, and show how it can be used in a reinforcement learning model, and a self-preserving agent model based on the free energy principle. We extend logical depth and conceptual jump size measures to stochastic domains, and elucidate their relations....
متن کاملkLog: A Language for Logical and Relational Learning with Kernels (Extended Abstract)
We introduce a novel approach to statistical relational learning; it is incorporated in the logical and relational learning language, kLog. While traditionally statistical relational learning combines probabilistic (graphical) models with logical and relational representations, kLog combines a kernelbased approach with expressive logical and relational representations. kLog allows users to spec...
متن کامل