Aries is a project aimed at exploring constraint-based techniques for automated planning and scheduling. It relies on an original implementation of constraint solver with optional variables and clause learning to which various automated planning problems can be submitted.
Previous research, especially with the FAPE solver, have shown that while obtaining state-of-the-art performance with constraint-based approaches to planning is possible, it requires pushing the inference capabilities of existing combinatorial solver beyond their current capabilities. The Aries project thus provides an innovative combinatorial solver that is built from the ground up by (1) mixing several techniques from constraint programming and automated reasoning, and (2) providing original representations and technologies relevant for automated planning:
- Finite domain CSP: at the core of a model are discrete variables with a finite domain denoting the set of possible values they can take.
- Literals represent expressions on the bounds of variables, for instance
(x <= 11)or(y > 10). This generalizes the literals in SAT solvers to non-boolean variables (whereas SMT or CSP solver typically maintain a correspondence table between such expressions and literals). - Various inference engines are provided in the solver:
- SAT based engine for disjunctive constraints (clauses in which at least one literal must be true), relying on unit clause propagation
- Difference Logic engine (aka STN), for propagating temporal constraints or general difference constraints between two variables. The difference logic engine notably supports forward checking (or theory propagation) and native reasoning on optional variables.
- General purpose CP engine for adding arbitrary constraints and the associated propagators.
- Explanation and Clause learning are supported by the various engine. When a conflict is detected during search, a new constraint will be inferred that prevents the solver of doing the same mistake.
- Optional variables: some variables can be optional their presence in the solution will be determined by the value of a literal. This allows eager reasoning and constraint propagation by decoupling the presence literal and the domain of the variable.
While the aries solver library is built with automated planning problems in mind, it remains a general purpose solver that can be used for other combinatorial problems. In particular, we provide solvers for SAT and jobshop problems as a thin wrapper around the library and can be used for testing or demonstration purposes.
Aries support problems in the PDDL and HDDL formats for specifying problems. A planning problem is then translated into an internal representation based on chronicles: data structures that specify the requirements ond effects of an action. Chronicles allow a rich temporal representation of an action and is especially useful for representing hierarchical problems where an abstract action can be decomposed into finer-grained ones.
THis representation allows for quite natural encoding of the planning problem into a constraint satisfaction problem that can be solved with our own combinatorial solver.
The current focus of the solver is on hierarchical planning which is especially well suited to represent various robotic and scheduling problems. Non-hierarchical problems are supported but do require more work to reach a state-of-the-art performance (areas of improvement notably include better symmetry breaking constraints and search heuristics).
First make sure you have rust installed on your machine.
Build relies on cargo can be triggered like so:
cargo build # debug non-optimized build
cargo build --release # optimized buildThis compiles artifacts and place them in the target/debug or target/release directory.
This repository contains several crates that provide various functionalities for automated problem solving. In a topological order :
core: Low-level representation of variables, domains, literals. It also provides the implementation for the state (collection of domains) supporting backtracking and explanations.model: Higher-level data structures and API to represent (typed) variables, expressions and their combination into constraint satisfaction problems.collections: Various collections that are focused on key-value collection where the key can be transformed into an array index.env_param: Utils to read parameters from environment variables.backtrack: Data structures for implementing trails that record event and allow undoing them. This crate provides a trail implementation meant for internal use in backtrackable structures and one that allows other entities to read the events that were pushed to the queue.solver: Implementation of combinatorial solvers with capabilities from SAT, SMT and CSP solvers. The solver provides an interface that accepts additional reasoners.reasoners: Specialized reasoners that provide inference capabilities to main solver. It currently includes a reasoner for difference logic and one experimental CP-like module.planning/planners: A crate that supports manipulating and solving AI Planning problems.examples: Several thin wrappers around the library to demonstrate and test its capabilities. Notably:sat: A thin wrapper around thesolvercrate that implements a sat executable that solves problems from CNF files.scheduler: A solver for several disjunctive scheduling problemsgg: A forward-search automated planner, in the spirit of YAHSP2.
While aries is primarily thought as a library it does come with some programs to allow testing and demonstration. Those include:
lcp: a plan-space planner for PDDL and HDDL, based on a compilation to constraint satisfaction problems.gg: a state-space planner for PDDL based on heuristic search with thehaddheuristic.aries-sat: SAT solver that mimic the minisat solver behaviorscheduler: solver for the jobshop and openshop problems
Source code of these executables can be found in the directory apps/src/bin. One can install an executable locally like so (example for gg):
cargo install --bin gg --path . # should be done in the "apps/" sub-crateThese can also be run directly with cargo. For instance for gg:
cargo run --release --bin gg -- <arguments>An overview of the concepts and algorithms at play in the aries project is provided as a mdbook in the doc/ folder.
This documentation is centered on the core part of the solver, dedicated to combinatorial problem solving with optional variables.
To build it, you should first install the mdBook command line tool. Then the following command should build the book and serve it locally (rebuilding on modifications). Please refer to the mdbook documentation for an overview of the other features.
mdbook serve doc/