We are pleased to announce the release of PRISM2.1 which is available for download at:
http://sato-www.cs.titech.ac.jp/prism/. This is a major update enhanced by three new inference methods (VT,VB-VT,MCMC) in addition to already available ones (EM,MAP-EM,VB-EM,DA-EM). PRISM is a logic-based probabilistic modeling language offering a declarative interface between users and machine learning tasks. As a probabilistic extension of Prolog, it is Turing-complete and covers a wide range of known models (BNs,HMMs,PCFGs,etc) and also unexplored models you define by PRISM programs. PRISM can reduce pains in probabilistic modeling by making available high-level logical expressions together with high level built-in predicates for machine learning tasks listed below. [1] Exact probability computation: efficient dynamic programming used [2] Sampling: forward sampling, MCMC (Metropolis-Hastings style) for Bayesian inference [3] Parameter learning: EM, DA-EM(deterministic annealing EM), MAP-EM, VT(Viterbi training, hard EM) [4] Approximate Bayesian inference: VB(variational Bayes)-EM, VB-VT [5] Viterbi inference: by parameters or by posterior distributions obtained from MCMC [6] Model score computation: BIC, Cheesman-Stutz score, VFE(variational free energy), log marginal likelihood via MCMC [7] Computing standard statistics You can immediately check these functionalities with many test programs that come with a self-contained user manual in the PRISM package. PRISM helps you identify and develop the best (Bayesian, non-Bayeian) model for your problem. With Best Regards, Taisuke Sato, Neng-Fa Zhou, Yoshitaka Kameya _______________________________________________ uai mailing list uai@ENGR.ORST.EDU https://secure.engr.oregonstate.edu/mailman/listinfo/uai
