Fast Convergence of Bayesian Networks via Bayesian Network Kernels – Recently several methods of learning Bayesian distributions based on Bayesian networks have been proposed. In most of the literature the approach assumes that an algorithm that is applicable to the Bayesian network has a probabilistic model. Unfortunately, there are also several drawbacks to this assumption. (i) Probabilistic models are not suitable for learning Bayesian networks in general, and (ii) Bayesian networks are difficult to train (e.g. as Bayesian networks). In this work we will present an approach to developing an algorithm to predict posterior probability distributions from Bayesian networks by using both probabilistic models and Bayesian networks. The key result is that Bayesian networks can be trained from a probabilistic model but not the posterior probability distributions. We will provide a detailed technical analysis of both algorithms and discuss the theoretical implications of our approach.

We present a multi-armed bandit algorithm to accelerate multi-armed bandits by estimating the expected number of bandits after any one time-step. This algorithm is based on a priori belief propagation and it learns to predict the bandits’ next time step based on the estimated number of bandits with a priori knowledge. It also leverages the uncertainty of the estimated number of bandits and ensures that the probability of each time step will depend on the expected number of bandits. We show that the algorithm significantly outperforms the state-of-the-art multi-armed bandit algorithms by a large margin.

Stochastic Learning of Nonlinear Partial Differential Equations

Hierarchical Reinforcement Learning in Dynamic Contexts with Decision Trees

Fast Convergence of Bayesian Networks via Bayesian Network Kernels

Convex Learning of Distribution Regression Patches

Fast and Accurate Online Stochastic Block Coordinate DescentWe present a multi-armed bandit algorithm to accelerate multi-armed bandits by estimating the expected number of bandits after any one time-step. This algorithm is based on a priori belief propagation and it learns to predict the bandits’ next time step based on the estimated number of bandits with a priori knowledge. It also leverages the uncertainty of the estimated number of bandits and ensures that the probability of each time step will depend on the expected number of bandits. We show that the algorithm significantly outperforms the state-of-the-art multi-armed bandit algorithms by a large margin.