Decisions.jl is an ecosystem for canonical representations of decision problems, from the most basic Markov decision processes, to rich and expressive multi-agent, semi-Markov, multi-objective extensions, unifying decision-, control-, and game-theoretic models.
Warning
Decisions.jl is under active development.
Decisions.jl represents decision problems using their underlying decision networks. A very broad and useful class of problems can be represented in this way.
Decisions.jl provides transformations to change problem classes with minimal disruption to the domain or solver implementation. Common transformations from adding agents to belief spacing - which usually require a frustrating amount of code rewriting in a bespoke framework - can be deployed to Decisions.jl problems in just a few lines. Using these transformations, it's easy to intuitively navigate the space of problem classes by iteratively adding or removing components.
Decisions.jl supports problem classes from the very common to the niche to the yet-unnamed. When possible, meaningful traits (like "partially observable" or "multiagent") relate these classes, forming a natural scaffold that can be traversed with transformations.
By leveraging Julia's just-in-time compilation, standard frameworking (sampling, rollouts, and so on) is automatically built for any decision problem that can be represented with a DDN. No more hand-writing simulator code for slight problem variations.
Decisions.jl is factored into three framework packages:
- DecisionNetworks.jl (closed beta) provides fundamental tools for the ecosystem: decision networks, conditional distributions, support spaces, and visualizations.
- DecisionProblems.jl (closed alpha) introduces objectives over decision networks and formal definitions of decision problems.
- DecisionSettings.jl (closed alpha) introduces real-world decision making scenarios, surgically defining concepts of agents, training loops, and environment interactions to permit truly exact comparisons between algorithms.
... and two implementation packages:
- DecisionDomains.jl provides implementations of common baseline decision making domains for benchmarking.
- DecisionAlgorithms.jl provides off-the-shelf implementations of classic decision-making algorithm.
If you're contributing to Decisions.jl, or you don't mind some unnecessary
dependencies, the package Decisions itself reexports all names from the packages listed
above, so you can just write using Decisions.
Decisions.jl is not (yet) a registered Julia package. To use it, clone the repo and do:
julia --project
] dev path_to_repo
(in your desired Julia environment) to add it in development mode. Of course, if you're working on Decisions.jl itself, no need to do this.
If you'd like a local copy of the Decisions.jl docs for development purposes, do:
cd ./docs
julia --project
include("make.jl")