AGGA

This is an implementation of AGGA a configurator for anytime portfolios.

The data (instances, scenarios, arguments, wrapper) used for our experiments can be found here

Installation

Install the requirements

pip install -r requirements.txt

To run configurations in parallel, we use ray. We require Ray version 2.3.1. By default, AGGA will make use of all available cores on a machine. To run it on a cluster, we provide a Slurm script that starts Ray before calling AGGA.

To construct the portfolios, Gurobi is needed.

Running AGGA

Search

Run AGGA by:

python agga/main.py --scenario_file <path_to_your_scenario> --arg_file <path_to_your_argument_file>

Arguments

Examples of how to run AGGA can be found here: here

Scenario

The scenario file should contain the following arguments:

• cutoff_time: Time the target algorithm is allowed to run
• instance_file: File containing the paths to the instances
• paramfile: Parameter file of the target algorithm in PCS format
• wallclock_limit: Time AGGA is allowed to run

Run target algorithm

To run the target algorithm, a wrapper is needed that, given the configuration, instance, and seed, returns a command string calling the target algorithm:

• wrapper_mod_name: Path to the wrapper file. The wrapper file will be imported by AGGA.
• wrapper_class_name: The class in wrapper_mod_name that should be called by AGGA to generate the command for the target algorithm.
• quality_match: Regex that signals AGGA that a solution can be found in the output line.
• quality_extract: Regex that extracts the solution quality from the previously matched line.
• memory_limit: Memory limit for a target algorithm run

Generell

• seed: Seed to be used by AGGA
• log_folder: Path to a folder where AGGA should write the logs and results
• num_cpu: Number of CPUs to be used
• localmode: If set to true, Ray will start in local mode, i.e., not look for a running Ray instance in the background
• termination_criterion [total_runtime, total_tournament_number]: Stopping criteria used by AGGA
• total_tournament_number [optional]: Number of AGGA iterations if termination_criterion is total_tournament_number
• use_ggapp: Create new configurations with the GGA++ model or by genetic engineering
• gga_rand_ratio: Percentage of random configurations to use with the GGA++ model
• capping [cost_sensitive, cap_opt, None]: Type of capping mechanism
• cap_model [rf, rp, xgb]: Type of capping models to be used with cost-sensitive capping
• cap_start: Time when the capping should start in seconds
• retrain_capping: Whether to retrain or not the cost-sensitive capping
• rf [optional]: Path to a JSON file containing the reference quality for the instances in the instance set

Instance set

• initial_instance_set_size: Size of the starting instance set used
• instance_increment_size: Number of instances that should be added in an iteration to the instance set
• target_reach [optional]: Approximation of the iteration by which the full instance set should be used
• time_instance_set_full: Time in seconds AGGA should use the full instance set. This may help with obtaining trajectories for all current configurations to be used in the portfolio construction

Portfolio Construction

After the search phase of AGGA, you can construct the final portfolios by calling:

python agga/portfolio_construction.py --agga_dir <directory_of_AGGA_run> --scenario_path <path_to_scenario_of_AGGA_run> --T <time_horizon_of_portfolio> --confs_portfolio <number_of_confs_to_include_in_portfolio> --confs_per_front <max_number_of_points_a_trajectory_can_contribute> --restrict_confs <restrict_number_of_confs_for_mip> --conf_restriction <number_of_confs_to_consider_in_mip> --gurobi_time_limit <timelimit_for_mip>

The portfolio will be saved in the directory of the AGGA run.