A list of ideas for future improvement of the package:

[ ] Add one or more new pairwise types to handle locally-adaptive pairwise association.
    Also create associated AL/ALR types (e.g. ALRadaptive).
[ ] Add MCML inference as described in Hughes, Haran, and Caragea (2011)
[ ] Add Bayesian inference as described in Hughes, Haran, and Caragea (2011)? 
    (maybe not-it's impractically slow...)
[ ] Think about how partially-observed graphs (missing responses) could be handled.
[ ] Implement Swendsen-Wang (either as a MC sampler or bounding chain-type perfect
    sampler) (see Huber, 2016 book).
[ ] Consider re-implementing parameters as arrays <: Real, instead of Float64, to allow use 
    of ForwardDiff.jl to get derivatives by automatic differentiation in optimization and Hessian
    calculation. Does it actually help optimization performance/accuracy/reliability? Any speed 
    trade-offs elsewhere due to use of generic type? Any numerical problems with existing 
    implementation that would justify the trouble of making the change?
[ ] Consider using StatsModels.DataFrameRegressionModel approach as in GLM for supplying predictors
    and responses as a DataFrame.
[ ] Consider changing FullPairwise to allow one parameter per edge *per observation*, for
    consistency with how FullUnary is defined. (currently every observation gets the same 
    Λ in FullPairwise)