Mercurial > hg > camir-aes2014

comparison toolboxes/FullBNT-1.0.7/docs/majorFeatures.html @ 0:e9a9cd732c1e tip

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
first hg version after svn
author wolffd
date Tue, 10 Feb 2015 15:05:51 +0000
parents
children
comparison
equal deleted inserted replaced
-1:000000000000 0:e9a9cd732c1e
1
2 <h2><a name="features">Major features</h2>
3 <ul>
4
5 <li> BNT supports many types of
6 <b>conditional probability distributions</b> (nodes),
7 and it is easy to add more.
8 <ul>
9 <li>Tabular (multinomial)
10 <li>Gaussian
11 <li>Softmax (logistic/ sigmoid)
12 <li>Multi-layer perceptron (neural network)
13 <li>Noisy-or
14 <li>Deterministic
15 </ul>
16 <p>
17
18 <li> BNT supports <b>decision and utility nodes</b>, as well as chance
19 nodes,
20 i.e., influence diagrams as well as Bayes nets.
21 <p>
22
23 <li> BNT supports static and dynamic BNs (useful for modelling dynamical systems
24 and sequence data).
25 <p>
26
27 <li> BNT supports many different <b>inference algorithms</b>,
28 and it is easy to add more.
29
30 <ul>
31 <li> Exact inference for static BNs:
32 <ul>
33 <li>junction tree
34 <li>variable elimination
35 <li>brute force enumeration (for discrete nets)
36 <li>linear algebra (for Gaussian nets)
37 <li>Pearl's algorithm (for polytrees)
38 <li>quickscore (for QMR)
39 </ul>
40
41 <p>
42 <li> Approximate inference for static BNs:
43 <ul>
44 <li>likelihood weighting
45 <li> Gibbs sampling
46 <li>loopy belief propagation
47 </ul>
48
49 <p>
50 <li> Exact inference for DBNs:
51 <ul>
52 <li>junction tree
53 <li>frontier algorithm
54 <li>forwards-backwards (for HMMs)
55 <li>Kalman-RTS (for LDSs)
56 </ul>
57
58 <p>
59 <li> Approximate inference for DBNs:
60 <ul>
61 <li>Boyen-Koller
62 <li>factored-frontier/loopy belief propagation
63 </ul>
64
65 </ul>
66 <p>
67
68 <li>
69 BNT supports several methods for <b>parameter learning</b>,
70 and it is easy to add more.
71 <ul>
72
73 <li> Batch MLE/MAP parameter learning using EM.
74 (Each node type has its own M method, e.g. softmax nodes use IRLS,<br>
75 and each inference engine has its own E method, so the code is fully modular.)
76
77 <li> Sequential/batch Bayesian parameter learning (for fully observed tabular nodes only).
78 </ul>
79
80
81 <p>
82 <li>
83 BNT supports several methods for <b>regularization</b>,
84 and it is easy to add more.
85 <ul>
86 <li> Any node can have its parameters clamped (made non-adjustable).
87 <li> Any set of compatible nodes can have their parameters tied (c.f.,
88 weight sharing in a neural net).
89 <li> Some node types (e.g., tabular) supports priors for MAP estimation.
90 <li> Gaussian covariance matrices can be declared full or diagonal, and can
91 be tied across states of their discrete parents (if any).
92 </ul>
93
94 <p>
95 <li>
96 BNT supports several methods for <b>structure learning</b>,
97 and it is easy to add more.
98 <ul>
99
100 <li> Bayesian structure learning,
101 using MCMC or local search (for fully observed tabular nodes only).
102
103 <li> Constraint-based structure learning (IC/PC and IC*/FCI).
104 </ul>
105
106
107 <p>
108 <li> The source code is extensively documented, object-oriented, and free, making it
109 an excellent tool for teaching, research and rapid prototyping.
110
111 </ul>
112
113