Accelerated Adaptive Markov Chain for Partition Function Computation
Ermon, Stefano, Gomes, Carla P., Sabharwal, Ashish, Selman, Bart
–Neural Information Processing Systems
We propose a novel Adaptive Markov Chain Monte Carlo algorithm to compute the partition function. In particular, we show how to accelerate a flat histogram sampling technique by significantly reducing the number of ``null moves'' in the chain, while maintaining asymptotic convergence properties. Our experiments show that our method converges quickly to highly accurate solutions on a range of benchmark instances, outperforming other state-of-the-art methods such as IJGP, TRW, and Gibbs sampling both in run-time and accuracy. We also show how obtaining a so-called density of states distribution allows for efficient weight learning in Markov Logic theories.
Neural Information Processing Systems
Dec-31-2011
- Country:
- North America > United States > Massachusetts (0.46)
- Genre:
- Research Report > Promising Solution (0.34)
- Technology: