Installation¶

Requirements¶

Julia 1.12 or later. The pinned Manifest.toml was resolved with Julia 1.12.6; continuous integration runs the test suite on 1.12 and 1.13.
A C toolchain is not required; all dependencies are pure Julia or ship binaries through Julia's artifact system.

Install¶

Not in the General registry

ARIS is not registered, so Pkg.add("ARIS") will not find it. Install from a clone.

git clone https://github.com/sisl/ARIS.jl
cd ARIS.jl
julia --project=. -e 'using Pkg; Pkg.instantiate()'

Pkg.instantiate() installs the exact versions recorded in Manifest.toml. That is the environment the paper's results were produced in, and it is what you want for reproduction. To work against freshly resolved dependencies instead, delete Manifest.toml before instantiating — and expect numerical results to move.

Check the install¶

julia --project=. -e 'using Pkg; Pkg.test()'

The suite covers the public API surface, the utility functions, the buffer, the conditional models, the sampling strategies, the estimator, and the baselines. It does not run the paper experiments.

A quick end-to-end check, which takes a couple of minutes:

julia --project=. examples/basic_usage.jl

Using it from another project¶

Add the clone as a development dependency:

using Pkg
Pkg.develop(path = "/path/to/ARIS.jl")

Then using ARIS as normal. Note that the example problems under examples/problems/ are not part of the package module — they are plain scripts you include, so that they can depend on things (plotting, driving environments) that the package itself does not.

Optional integrations¶

Two integrations are gated off by default and are not used by any paper experiment:

gate	environment variable	brings in
normalizing flows	`CGV_ENABLE_FLOW=true`	`PyCall`, and Python `torch` + `normflows`
policy / MDP proposals	`CGV_ENABLE_POLICY=true`	the `RMDP` and conditional-policy code paths

These are read at module load, so set them before using ARIS. Their dependencies are not in Project.toml; you must provide them in the active environment yourself. Leave both unset unless you specifically need them.

Reproducing the paper¶

Experiment 4 uses a separate environment, because it pulls a pinned Git dependency:

julia --project=paper/envs/idm -e 'using Pkg; Pkg.instantiate()'

The figure and metric scripts are Python and need numpy and matplotlib. See Reproducing experiments.