Towards BayesNets 2.0

[tawheeler]

We dropped support for Julia 0.4, so the next release will need to increment the major tick. Thus, BayesNets v2.0.0

Before I tag a new release I would like to:

• merge in PR Inference #42
• merge in PR Staged changes for final project in forked repo #43
• verify docs are up to date
• make sure all samplers use AbstractSampler
• resolve problem of having two Gibbs samplers
• make inference more consistent, probably with an AbstractInference type and associated function
• resolve Likelihood weighted inference does not return a table over all possible assignments #49
• resolve Loopy Belief inference is only implemented for a single node. #48

Does anyone have any additional issues to post? Things we need to clear up before moving forward?

[mykelk]

Looks good to me!

[tawheeler]

We now have a common interface for samplers:

abstract BayesNetSampler
Base.rand!(a::Assignment, bn::BayesNet, sampler::BayesNetSampler)

The only function new samplers need to implement is rand!

From that we provide:

Base.rand(bn::BayesNet, sampler::BayesNetSampler)
Base.rand(bn::BayesNet, sampler::BayesNetSampler, nsamples::Integer)

rand(bn) defaults to DirectSampler
rand(bn, N) also
rand(bn, evidence) defaults to RejectionSampler
rand(bn, N, evidence) also

[tawheeler]

I introduced a new common interface for inference:

"""
Abstract type for probability inference
"""
abstract InferenceMethod

"""
Infer p(query|evidence)
 - inference on a DiscreteBayesNet will always return a DataFrame factor over the evidence variables
"""
infer(im::InferenceMethod, bn::BayesNet, query::Vector{NodeName}; evidence::Assignment=Assignment()) = error("infer not implemented for $(typeof(im))")
infer(im::InferenceMethod, bn::BayesNet, query::NodeName; evidence::Assignment=Assignment()) = infer(im, bn, [query]; evidence=evidence)

Ideally it would return a distribution over the queried variables, ie a multivariable distribution. That isn't really possible right now, so for discrete variables we are using factors (which are basically the same thing) and for continuous / hybrid we can't really do anything.

[tawheeler]

New Inference branch. Incorporated updated (the excellent) inference methods from @hamzaelsaawy