PRISM1.8 Released

From: Neng-Fa Zhou (nzhou_at_acm.org)
Date: 12/03/04 

Date: Fri, 03 Dec 2004 01:51:06 GMT

We are pleased to announce the release of PRISM1.8, which is available for
download at:

 http://sato-www.cs.titech.ac.jp/prism/.

PRISM is a logic-based probabilistic language and is easy to learn and use
for users who are familiar with Prolog. The most notable feature of PRISM
is that it allows the user to define random switches and use them to make
probabilistic choices. The probability distributions of switches can be
learned automatically from samples. PRISM is suited for building complex
systems that involve both symbolic and probabilistic elements such as
discrete hidden Markov models, stochastic string/graph grammars, game
analysis, data mining, performance tuning and bio-sequence analysis.

PRISM1.8 improves upon the previous version 1.7 in the following aspects:

・瘢雹 Significant improvement of learning speed. For instance, learning on a
real context free grammar with 860 productions using a corpus of 10000
sentences, which was impossible before with version 1.7, can be completed in
only 10 minutes on a PC.

・瘢雹 While early versions accepted only failure-free and negation-free
programs, the new version is able to handle failure and negation. Negation
is compiled away automatically when programs are loaded. As a result, we can
describe models with failure constraints such as the agreement in number
between subjects and verbs. In addition, learning can be conducted with
negative samples.

PRISM, in general, enjoys the following features:

 (1) The user can use programs to define distributions over terms and atoms.

Mathematically a PRISM program is a formalism, which defines a probability
measure over the set of possible Herbrand interpretations. The distributions
are derived and computed from the defined measure. There are no restrictions
on programs, e.g., programs are not required to be range-restricted.

(2) Parameters in a program are learnable automatically from examples.

A PRISM program contains statistical parameters that reflect the statistical
properties of the model. They can be automatically estimated from examples
by ML (Maximum Likelihood) estimation performed by a built-in EM learning
routine.

(3) Probabilities are computed efficiently in a dynamic programming manner.

PRISM uses "explanation graphs" to compute probabilities and learn
parameters, where solutions are shared as in dynamic programming. PRISM is
implemented on up of B-Prolog and tabled search of B-Prolog is used to
construct explanation graphs.

(4) PRISM is a high level yet efficient modeling language.

Popular symbolic-statistical models such as hidden Markov models,
probabilistic context free grammars and Bayesian nets can be described in
PRISM in a very compact way. Parameter learning in PRISM can be done as
efficiently as those by specialized EM algorithms such as the Baum-Welch
algorithm. In addition, PRISM can be used to model certain phenomena that
are hard to model using the specialized statistical tools.

PRISM is sustainable to relatively large sets of data and would be of
interest to anyone who would like to challenge statistical modeling of
complex phenomena.

With best regards,

Taisuke Sato and Neng-Fa Zhou