Prost is a probabilistic planning system that repeatedly computes which action to take in the current state, executes that action by interacting with rddlsim and updates the current state according to the outcome. The input language of Prost is RDDL.
Make sure you meet the following requirements before installing Prost:
- Linux operating system
- git is installed to clone the repository
- flex 2.6.4 or higher and bison 3.0.4 or higher are installed to parse the input
- cmake and a c++ compiler (we recommend gcc) are installed to compile the planner
- the BDD library BuDDy is installed
- the theorem prover z3 is installed
On Linux, you can install all dependencies except for z3 with the command
sudo apt install git g++ cmake bison flex libbdd-dev
Installation information for z3 can be found here. Without any guarantee for success, you can try the following steps to install z3:
git clone git@github.com:Z3Prover/z3.git /path/to/z3-repo(clone z3 repo)cd /path/to/z3-repo(switch to directory)python3 scripts/mk_make.py --prefix=/path/to/z3-install(configure and pass installation location)cd build(switch to build directory)make(build)make install(install)
If BuDDy or z3 is not found automatically, you can set the environment variables
$BDD_ROOT and Z3_ROOT, which point to your BuDDY and z3 installation, respectively.
Prost executes the policy it computes by interacting with the rddlsim server. We refer to the github repository of rddlsim for the latest installation information. Without any guarantee for success, you can try the following steps to install rddlsim:
sudo apt install default-jre(install Java runtime environment)git clone https://github.com/ssanner/rddlsim /path/to/rddlsim(clone rddlsim repo)cd /path/to/rddlsim(switch to directory)./compile(compile rddlsim)
The Prost repo contains a script to run rddlsim
('testbed/run-server.py'). It requires that you set an environment
variable $RDDLSIM_ROOT that points to /path/to/rddlsim, e.g. by adding
'export RDDLSIM_ROOT=/path/to/rddlsim'
to the file ~/.bashrc. You don't have to use this script to run rddlsim, but we assume in the following that you do.
Installation of Prost is simple and can be done in three steps:
git clone https://github.com/prost-planner/prost /path/to/prost(clone Prost repo)cd /path/to/prost(switch to directory)./build.py(compile the planner in release mode)
If you need to compile the planner in debug mode, replace the last
command with ./build.py --debug.
Before you can run Prost, you have to start the rddlsim server. One way to do so is by using the provided script that can be found in the testbed folder. For instance, running
./run-server.py -b benchmarks/elevators-2011/
from the testbed directory loads all instances of the elevators domain
of IPC 2011 into rddlsim (run ./run-server.py --help for a full list
of available options).
Now you can start Prost with the prost.py script that can be found in
the root directory of the repo. For instance, to start planning for the
first elevators instance of IPC 2011 with the Prost version from IPC
2014, run
./prost.py elevators_inst_mdp__1 "[Prost -s 1 -se [IPC2014]]"
Note that there are many Prost configuration implemented. Run prost.py
without additional parameters for an overview of all available solvers
and their options.
We recommend to perform experiments with the Prost planner by using Prost Lab, an experiment package that is based on the Lab toolkit. Information on how to set up Prost Lab can be found here.
If you want to compare your planner to Prost, please make sure to
compile Prost in release mode and to use an appropriate search engine for
Prost. If in doubt, run both the Prost-2011 and Prost-2014 planners by
using -se [IPC2011] and -se [IPC2014] in the call to prost.py. If
you don't have the time to use both configurations, we recommend to go
with Prost-2014.
Please cite the Prost planner as well as the Prost-2011 version of the planner with
@InProceedings{keller-eyerich-icaps2012,
author = "Thomas Keller and Patrick Eyerich",
title = "{PROST}: Probabilistic Planning Based on {UCT}",
booktitle = "Proceedings of the Twenty-Second International Conference on
Automated Planning and Scheduling (ICAPS 2012)",
year = "2012",
publisher = "AAAI Press"
pages = "119--127"
}
and the Trial-based Heuristic Tree Search framework or the Prost-2014 version of Prost with either
@InProceedings{keller-helmert-icaps2013,
author = "Thomas Keller and Malte Helmert",
title = "Trial-based Heuristic Tree Search for Finite Horizon {MDPs}",
booktitle = "Proceedings of the Twenty-Third International Conference on
Automated Planning and Scheduling (ICAPS 2013)",
year = "2013",
publisher = "AAAI Press"
pages = "135--143"
}
or
@PhdThesis{keller-phd2015,
author = "Thomas Keller",
title = "Anytime Optimal {MDP} Planning with Trial-based Heuristic Tree Search",
school = "University of Freiburg",
year = "2015"
}
If you have any questions, feel free to contact one or all of the currently active code maintainers:
There are quite a few additional people who have contributed to Prost since the planner's development was started in 2010 (these are ordered alphabetically):
- Augusto Blaas Correa
- Dorde Relic
- Florent Teichteil-Koenigsbuch
- Guillem Frances
- Jendrik Seipp
- Joris Scharpff
- Max Grüner
- Patrick Eyerich
- Scott Sanner
There are also some code snippets from others that have found their way into the code base, which we'd like to give credit for:
- Sungwook Yoon (ipc client)
- René Nyffenegger (base64 encoding and decoding)
- Python Software Foundation (code to hash states)
The Prost planner is released under the MIT license with the exception of the following files:
- src/search/ipc_client.{h,cc}
- src/search/utils/strxml.{h,cc}
- src/search/utils/base64.{h,cc}
- src/search/utils/hash.{h,cc}
All of these files contain code from other sources that has been released under different licenses. The changes we made to the source code are provided 'as-is', without any express or implied warranty. In no event will the Prost contributors be held liable for any damages arising from the use of this software.