I have implemented the  Continuous restricted Boltzmann machine based on the original publication:

H. Chen and A.F. Murray. Continuous restricted Boltzmann machine with an implementable training algorithm , IEE Proc., Vis. Image Process. 150, 153 (2003).

The implementation is done using Octave, and the code is attached in the web page. I have reproduced the results for the artificial data as mentioned in the paper. I am currently experimenting CRBM with reinforcement learning, specially with actor-critique (Cacla) to learn the balancing of a Humanoid robot in continuous space, and actions. 

The following figures show the results for the learned distribution, and the traces of the noise parameters. The data set is created from sampling from two multivariate Gaussian distributions.

Noise parameter traces:

I have further tested the CRBM for non-Gaussian distributions as we can not assume that the robot sensor measurements are superimposed with Gaussian noise (please correct me If I am wrong). In order to model non-Gaussian distributions, we may have to fix the input layer noise parameters, and let the hidden layer noise parameters to be learned. This mechanism is used in:

Tong Boon Tang, Alan F. Murray: Adaptive sensor modelling and classification using a continuous restricted Boltzmann machine (CRBM). Neurocomputing 70(7-9): 1198-1206 (2007).

Non-Gaussian data set, and construction of the distribution:
Traces of the hidden layer noise parameters:

Saminda Abeyruwan,
Feb 2, 2012, 12:29 AM