Projektbeschreibung
Von datenbasierter zu modellbasierter künstlicher Intelligenz: Repräsentationenlernen für die Planung
Zwei Hauptforschungsthemen im Bereich der künstlichen Intelligenz (KI) befassen sich mit der Entwicklung datenbasiert lernender KI, die aus Erfahrungen und Daten auf Verhaltensweisen und Funktionen schließen können, und modellbasiert lernender KI, die mit unlösbaren Modellen wie SAT, klassischer Planung und Bayes‘schen Netzen umgehen können. Lernende KI-Systeme, insbesondere tief lernende KI, haben beachtliche Erfolge erzielt, führen jedoch zu unflexiblen Black-Boxen. Löser hingegen benötigen Modelle, die von Hand schwer zu erstellen sind. Das EU-finanzierte Projekt RLeap verfolgt das Ziel, eine Integration von beiden im Kontext der Planung zu erreichen, indem das Problem des Erlernens symbolischer Darstellungen erster Ordnung allein mithilfe unverfälschter Wahrnehmungen angegangen und gelöst wird. Das Projekt kann die Art und Weise, wie allgemeine, erklärbare und vertrauenswürdige KI verstanden und umgesetzt wird, grundlegend verändern.
Ziel
Two of the main research threads in AI revolve around the development of data-based learners capable of inferring behavior and functions from experience and data, and model-based solvers capable of tackling well-defined but intractable models like SAT, classical planning, and Bayesian networks. Learners, and in particular deep learners, have achieved considerable success but result in black boxes that do not have the flexibility, transparency, and generality of their model-based counterparts. Solvers, on the other hand, require models which are hard to build by hand. RLeap is aimed at achieving an integration of learners and solvers in the context of planning by addressing and solving the problem of learning first-order planning representations from raw perceptions alone without using any prior symbolic knowledge. The ability to construct first-order symbolic representations and using them for expressing, communicating, achieving, and recognizing goals is a main component of human intelligence and a fundamental, open research problem in AI. The success of RLeap requires the development of radically new ideas and methods that will build on those of a number of related areas that include planning, learning, knowledge representation, combinatorial optimization and SAT. The approach to be pursued is based on a clear separation between learning the symbolic representations themselves, that is cast as a combinatorial problem, and learning the interpretations of those representations, that is cast as a supervised learning problem from targets obtained from the first part. RLeap will address both problems, not just in the planning setting but in the generalized planning setting as well where plans are general strategies. The project can make a significant difference in how general, explainable, and trustworthy AI can be understood and achieved. The PI has made key contribution to the main themes of the project that make him uniquely qualified to carry it forward.
Wissenschaftliches Gebiet
Schlüsselbegriffe
Programm/Programme
Thema/Themen
Finanzierungsplan
ERC-ADG - Advanced GrantGastgebende Einrichtung
52062 Aachen
Deutschland