Bayesian Optimization with Automatic Prior Selection for Data-Efficient Direct Policy Search

Meta-repo for code implementing the MLEI acquisition function

Paper: "Bayesian Optimization with Automatic Prior Selection for Data-Efficient Direct Policy Search", submitted to the International Conference on Robotics and Automation (ICRA) 2018.

Authors: Rémi Pautrat, Konstantinos Chatzilygeroudis, and Jean-Baptiste Mouret.

How to use it

How to properly clone this repo

git clone --recursive https://github.com/resibots/pautrat_2018_mlei.git

Dependencies

• Boost: C++ Template Libraries
• Eigen: Linear Algebra C++ Library
• realpath: sudo apt-get install realpath

How to easily compile everything

Important: Make sure you have installed all the dependencies of each repo. Otherwise the build will fail.

From the root of this repo run:

sh compile_all.sh

Note: You have to set the two environment variables RESIBOTS_DIR and LD_LIBRARY_PATH before each use of this code, so you may want to add the two export commands to your .bashrc file:

export RESIBOTS_DIR=[PATH_OF_YOUR_ROOT_DIRECTORY]/install

export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:[PATH_OF_YOUR_ROOT_DIRECTORY]/install/dart_path/lib

How to run MAP-Elites experiments to generate priors

cd sferes2/build/exp/map_elites_hexapod/
./hexa_duty

The best results were obtained with a 6D descriptor, the body orientation, but other descriptors are available in hexapod_simu/hexapod_dart/include/hexapod_dart/descriptors.hpp.

How to use Bayesian optimization with selection of priors in a limbo experiment

cd limbo/build/exp/bo_mlei
./bo_mlei_flat_ground [directory_containing_your_priors] [options]

bo_mlei_flat_ground launches the simulation on flat ground, but you can use the other environments as well with bo_mlei_easy_stairs, bo_mlei_medium_stairs and bo_mlei_hard_stairs. Use bo_mlei_robot to use it on a real robot.

Several options are available:

• -l: id of the leg you want to remove (-1: no broken leg, 0 and 5: left and right rear legs, 1 and 4: left and right middle legs, 2 and 3: left and right front legs) (default: -1).
• -s: method used of prior selection (0: random, 1: constant prior, 2: MLEI, default: MLEI).
• -t : number of types of different priors that you use (default: 4, e.g. for flat ground, easy stairs, medium stairs and hard stairs priors).
• -p : number of priors used for each type of prior (default: 15).
• -n : number of iterations for Bayesian optimization (default: 10).
• -c : controller of the robot given by its 54 parameters (only for replaying one gait, there is no learning here).

How to easily clean everything

From the root of this repo run:

sh clear_all.sh

LICENSE

CeCILL