Usage

Necessary Provisions

After you installed rtac, you need to provide data for the AC scenario to be complete.

• Target algorithm.

• A text file containing a defitionion of the scenario.

  • Hyperparameters that you want to adjust.

  • A dotted module path to the target algorithm wrapper.

  • The class name of the wrapper

  • An absolute path to the instance feature generator if you want to use instance features.

  • The class name of the feature generator.

  • An absolute path to a csv file containing instance features if you precomputed them and the instance feature generator only reads them in.

  • A text file with instance features to pre-train CPPL if you use CPPL.

  • An absolute path to a json or pcs file defining the parameter space of the target algorithm.

• Problem instances.

• Absolute paths to the problem instances to be solved.

You can read in the scenario with scenario = rtac.utils.read_io.read_args(‘path_to_scenario_file’), initialize the RAC with rtac = rtac.rtac.rtac_factory(scenario) and solve instances with rtac.solve_instance(instance).

Hyperparameters

Below, you will find a table with hyperparameters accepted by the RAC methods in RTAC. You can set them within the scenario file or in the command line.

Hyperparameters for RTAC

Name	Type	Default	Description
–verbosity	int	0	Verbosity level. [0,1,2]
–number_cores	int	1	Number of cores to be used in parallel.
–wrapper	str	No wrapper chosen!	Module of Python wrapper for the algorithm.
–wrapper_name	str	No wrapper class name.	Name of the wrapper class in the wrapper module.
–feature_gen_name	str	No feature generator class name.	Name of the feature generator class in the feature generator module.
–feature_path	str	No feature directory path.	Path to the directory with feature files, if existing.
–timeout	int	300	Stop solving single instance after (int) seconds. [300]
–runtimePAR	int	1	Multiply –timeout by –runtimePAR if instance not solved. [1]
–contenders	int	30	The number of contenders in the pool. [30]
–keeptop	int	4	Number of top contenders automatically in tournament (ReACTR/ReACTR++).
–chance	int	25	Chance to replace gene randomly in ReACTR/ReACTR++. [0–100]
–mutate	int	10	Mutation chance in crossover in ReACTR/ReACTR++. [0–100]
–kill	float	5	Kill contenders with higher variance in ReACTR/ReACTR++.
–feature_gen	str	“”	Python wrapper for computing instance features.
–instance_pre_train	str	False	Path to pre-training feature file for the bandit.
–nc_pca_f	int	3	Number of PCA dimensions for instance features.
–nc_pca_p	int	5	Number of PCA dimensions for parameter features.
–jfm	str	polynomial	Joined feature map mode for CPPL.
–omega	float	1.0	Omega parameter for CPPL.
–gamma	float	1	Gamma parameter for CPPL.
–alpha	float	0.2	Alpha parameter for CPPL.
–epsilon	float	0.9	Epsilon for epsilon-greedy selection (0.0 = greedy).
–kappa	float	1.0	Weight on confidence in pairwise comparison (CPPL).
–gen_mult	int	2	Generation multiplier before CPPL insertion.
–ac	int	1	Algorithm Configuration method [1=ReACTR, 2=ReACTR++, etc.].
–paramlimit	float	100000	Parameter value limit for normalization in CPPL.
–win_bonus	float	0.2	Boost for recent CPPL winner skill.
–win_decay	float	0.8	Decay factor for CPPL skill estimates.
–recent_winner_boost	float	3.0	Absolute skill boost for recent CPPL winner.
–forgetting_factor	float	0.98	Forgetting factor for old observations (theta_bar).
–obs_noise	float	1.0	Assumed observation noise variance in CPPL.
–cppl_reward	float	1.0	Reward for winning arm in CPPL bandit model.
–cpplt	float	2	Reset value for t in CPPL when pool changes.
–log_folder	str	logs	Directory for logging.
–param_file	str	No parameter file given!	Path to PCS or RTAC JSON parameter file.
–resume	bool	False	Resume RTAC from logged configuration state.
–baselineperf	bool	False	Run only default parameterizations.
–pws	bool	False	Insert default configuration if flag set.
–objective_min	bool	False	Optimize for objective value minimization.
–experimental	bool	False	Use tournament 0 logs for experiment.
–online_instance_train	bool	False	Enable continuous PCA fitting on incoming instances.
–epsilon_greedy	bool	False	Enable epsilon greedy selection.
–isolate_bandit	bool	False	Run bandit in child process to avoid resource conflict.
–gray_box	bool	False	Enable gray-box RAC.
–gb_read_time	float	0.1	Frequency to check gray-box output [seconds].
–nr_gb_feats	int	2	Number of gray-box features used.

What To Use

What you can use depends on the availability of features in your setting and if you would like to include them. In the following, a table shows the possible RAC method use based on the feature availability.

Features	ReACTR	CPPL	ReACTR++	Gray-box ReACTR(++)	Gray-box CPPL
Problem Instance	✗	✓	✗	✗	✓
Runtime Output	✗	✗	✓	✓	✓