Anglican is tightly integrated with Clojure. In order to learn to program in the Anglican language you should probably invest some time to learn Clojure and you must install leiningen if you have not already. Optionally you may wish to familiarize yourself with Gorilla REPL.

Installing and Getting Started Guide

Full installation instructions and a getting started guide are available here.

Installing everything takes about 10 minutes. Learning Anglican a few minutes if you know Clojure. A few hours if you know functional programming – a bit longer if you are entirely unfamiliar with functional programming.

Using Anglican

An Anglican query can be either run as a standalone program, producing the inference results on the standard output, run via a web-based read eval print loop (REPL) notebook server (called Gorilla REPL), or invoked programmatically from Clojure code.

The anglican-user repository provides the environment and program templates for these styles of usage.

Calling Anglican from Clojure

The anglican-user repository contains a leiningen project.clj with the suggested dependencies and namespace declarations. Anglican is also available on Clojars.

An Anglican query can also be invoked programmatically from Clojure code, using the doquery macro, with the following syntax:

(doquery algorithm query value & options)

where algorithm is the keyword specifying the algorithm (:lmh, :pgibbs, :pcascade etc.), value is the initial value passed to the query (nil if the query does not expect arguments), and options are Clojure clojure keyword arguments, specific to each inference algorithm. doquery returns a lazy sequence of samples — objects encapsulating a map of predicts and the log probability of each sample.

The example worksheets provide many examples of how do use doquery. The language tab explains how to construct the Anglican programs which are the essence of query’s.

Browser-based REPL

Gorilla REPL provides a convenient environment for processing and visualization of inference results. worksheets/template.clj is supplied as a starting point for an Anglican worksheet. Running

lein gorilla

in the root of anglican-user starts a local http server which can be connected to via a local browser.

Running Standalone Programs

Standalone Anglican programs can be run from the command line or in the REPL. The template for a standalone program, which you can just copy and modify for your needs, is programs/template.clj. The command line syntax is

lein run namespace [program] [option ...]

from the command line, or:

(m! namespace [program] [option ...])

in the REPL, where namespace is the namespace containing the Anglican program to run, for example:

bash$ lein run branching -a pgibbs -n 100 \
           -o ":number-of-particles 50"

anglican.core=> (m! branching -a pgibbs -n 100
                  -o ":number-of-particles 50")

program is the first argument of defquery. The namespace may contain multiple programs. If program is omitted, it defaults to the last component of the namespace (hmm for anglican.hmm, logi for anglican.logi).

Options:

-a, --inference-algorithm NAME   :lmh       Inference algorithm
-b, --burn N                     0          Skip first N samples
-d, --debug                                 Print debugging information
-f, --output-format FORMAT       :anglican  Output format
-n, --number-of-samples N                   Output predicts for N samples
-o, --algorithm-options OPTIONS  []         Algorithm options
-t, --thin N                     1          Retain each Nth sample
-v, --value V                               Initial value to pass to the program
-w, --warmup FLAG                true       Pre-evaluate the program
-h, --help                                  Print usage summary and exit

Redirecting Input and Output

Inference output is normally redirected to a file for post-processing. On the command line, > can be used:

bash$ lein run branching -a pgibbs -n 100 \
           -o ":number-of-particles 50" > branching.pgibbs

In the REPL, a macro redir is provided. The syntax is:

(redir [:in "input-file-name" :out "output-file-name"]
  actions ...)

Either :in or :out (or both) can be omitted. if the output file name begins with +, + is removed and the output is appended to the file. In the REPL, the above command example can be run as

anglican.core=> (redir [:out "branching.pgibbs"]
                  (m! branching -a pgibbs -n 100
                      -o ":number-of-particles 50"))

Functions freqs (frequency table for every integer-valued or symbolic predict) and meansd (mean and standard deviation for each predict) can be used in the REPL to quickly assess the results:

 anglican.core=> (redir [:in "branching.pgibbs"] (freqs))
 r, 0, 0.0260000, -3.64966
 r, 1, 0.0930000, -2.37516
 r, 2, 0.0740000, -2.60369
 r, 4, 0.00100000, -6.90776
 r, 5, 0.340000, -1.07881
 r, 6, 0.244000, -1.41059
 r, 7, 0.133000, -2.01741
 r, 8, 0.0590000, -2.83022
 r, 9, 0.0170000, -4.07454
 r, 10, 0.00700000, -4.96185
 r, 11, 0.00500000, -5.29832
 r, 12, 0.00100000, -6.90776