Home Research Research Group Publications Teaching Service Data/Software Contact

Roni Khardon
Department of Computer Science
Indiana University, Bloomington

Research Interests and Projects: My interests are in the areas of Machine Learning and Data Mining, Artificial Intelligence, and Efficient Algorithms and my research explores theoretical questions, empirical questions, and applications. Most recently my work has focused on problems in two broad areas: Theory, Algorithms and Applications of Graphical models and Stochastic Planning (aka Reinforcement Learning) and its Relation to Probabilistic Inference.

Research Topic: Planning and Inference

This work is partly supported by NSF grant 2002393

Stochastic planning is the basic model for agents that act in their environment in order to maximize long term rewards. This is often captured mathematically using Markov decision processes (MDP) and the problem is often studied as Reinforcement Learning when the agent does not have a model of the environment to work with.

Our work has focused on scalabity of solutions for such problems, especially in discrete spaces and especially when the states and actions are high dimensional so that the optimization problem is hard. Past contributions include: learning for planning, symbolic dynamic programming in structured spaces (both propositional and relational). Our recent work explores approaches that take advantage of connections between stochastic planning and inference in probabilistic graphical models. The research combines theoreical analysis, new algorithms, and building systems. Our SOGBOFA system has recently participated in the International Planning Competition 2018 and was runner-up in the competition.

Links to systems and papers: (see publications)

Research Topic: Theory, Algorithms and Applications of Graphical Models

This work is partly supported by NSF grant 1906694.

Graphical Models is a general paradigm for AI and machine leanring approaches that are based on probabilistic models. Our work is done in the context of expressive Bayesian probabilistic models, developing inference algorithms for them, developing a learning theory that explains why these algorithms work and applying them in interesting applications. Recent applications include land-cover clustering and classification, analysis of time series from Astronomy, and predicting contamination level in environmental engineering.

Our recent work has focused on two aspects: developing general algorithms that are applicable to many models and develping a theory that explains why and when Bayesian machine learning algorithms work well. Our work introduced novel fixed-point algorithms that improve convergence speed of variational inference methods. Our theoretical results provide distribution-free guarantees on the risk of approximate Bayesian inference algorithms, specifically justifying the use of variational inference on some problems. The goal is to understand which approximations work well and under what conditions and to develop effective algorithms implementing successful approximations. Recent models include constrained clustering, multi-task learning, sparse Gaussian processes, mixture of expert models for label discretization, matrix facorization and topic models.

Links to systems and papers: (see publications)