Metropolis extension

[nossleinad]

Generalize metropolis_sample by introducing metropolis_sample(update!::Function, -||-)

Now, we can customize how a "metropolis step" looks, typically with the do-syntax (thanks @AntonOresten for the tip). Omitting update! is equivalent with the previous behaviour. This enables us to, for instance, fix branch lengths and only sample over topologies with nni_update!. We can also add parameters to the chain as exemplified below.

Example usage

From Phylotrajectories.jl:

trees, LLs = metropolis_sample(
        newt,
        [bm_model],
        model.n_samples,
        burn_in = model.burn_in,
        sample_interval = model.sample_interval,
        collect_LLs = true,
    ) do tree, models
        sample_leafs!(temp_messages, tree, x -> models, model.frequency_sampler)
        for i = 1:model.consecutive_root_samples
            sample_root_distribution!(temp_messages[1], tree, model.root_distribution_sampler)
        end
        nni_update!(softmax_sampler, tree, x -> models)
        branchlength_update!(model.branchlength_sampler, tree, x -> models)
    end

this extension called for introducing the

metropolis_step interface

metropolis_step(LL::Function, modifier, curr_value)

Does a standard metropolis step in an MCMC, i.e. proposes a candidate symmetrically and returns the next
state in the chain, decided by the candidate being rejected or not.

The interface for metropolis_step is as follows
You need a MySampler <: Any to implement

• proposal(modifier::MySampler, curr_value)
• log_prior(modifier::MySampler, x)
• apply_decision(modifier::MySampler, accept::Bool)

An example is given for BranchlengthSampler

Add branchlength_update! & nni_update!

Update can be either an optim or a sampling, determined by the first argument passed in. We still export branchlength_optim! and nni_optim! and they behave identically as before. This separation is made to avoid confusion when we want *_optim! to sample.