Statistical Relational AI at KI-2018

The 41st German Conference on Artificial Intelligence

September 24-28th, 2018, Berlin, Germany

Tanya
Braun is a research associate at Universität zu
Lübeck,

Kristian Kersting is a professor
for Machine Learning at
Darmstadt University, and

Ralf Möller is a professor for
Information Systems at Universität zu Lübeck

Probabilistic relational modeling for artificial intelligence (also called statistical relational AI) has attracted the interest of a huge number of researchers in the last ten years. Assuming a background in probability theory as well as in basics about Bayesian or Markovian networks, we introduce specific probabilistic relational representation languages, exact inference algorithms, and learning techniques. Using a tele-healthcare application example we demonstrate achievements and challenges.

- Introduction (Kristian)
- Probabilistic relational models (PRMs)
- StarAI
- Systems AI

- Probabilistic relational models: languages and applications (Ralf)
- Landscape overview, semantics (grounded-distributional, maximum entropy principle)
- Inference problems (margin queries with and without evidence, filtering, smoothing, most-probable explanation, maximum a posterior)
- Lifted inference for grounded semantics: central ideas
- Learning parameters and structure of PRMs, basics: maximum likelihood, expectation maximization, first-order inductive learning: where PRM learning started
- Reusing ontologies for PRMs, computing ontologies for the job from PRMs
- Actions and utilties, simple decisions: maximum expected utility, sequential decisions: partially observable first-order Markov decision problems

- Specific topics on inference and learning
- Exact lifted inference (Tanya)
- Exact lifted dynamic inference (Tanya)
- Approximate lifted inference (Kristian)
- Relational reinforcement learning (Kristian)

- Summary (Kristian)