Inference Methods

From Clojure, the syntax for calling an inference algorithm on a query is (doquery :method foo args & method-opts). For example, we could run the smc algorithm with :number-of-particles set to 100 on a program foo defined by (defquery foo [x] ...) as follows

(doquery :smc
  foo [xval]
  :number-of-particles 100)

When a query object is passed to doquery or conditional the user has a choice of what kind of inference should be used to perform the characterization of the denoted conditional distribution. In general we recommend rmh if all observes occur at the end of the program, smc with as many particles as can fit in memory if observes occur with the program, and ipmcmc for programs with internal observes that are too large for smc to solve well.

CRITICAL: the default arguments supplied for all inference methods are the minimum required. This design make interactive debugging easier but nearly absolutely ensures poor inference performance unless non-default parameter values are chosen. For instance smc performance will only get better with more particles; ultimately how many particles to use should be chosen so as to maximize the joint criteria of how long you want to wait and how good you want your results to be.

Anglican provides 4 general categories of inference methods and several instances of each:

• Importance sampling (IS),
• Markov chain Monte Carlo (MCMC),
• Particle Markov chain Monte Carlo (PMCMC),
• Maximum a posteriori (MAP).

The current list of algorithms that are included with Anglican is

Method-specific Options

Some algorithms have optional arguments, which can be specified with (doquery :method prog args :option value). These are