Meta Parameter Settings

Intro

On top of the normal settings file that dictates the settings for the learning simulation, there is another file that lists the settings for a Meta simulation. These simulations usually involve samples of learning simulations and potentially changing some learning settings for each batch of samples. Also due to the nature of the meta settings they are often computer specific. For example, there is an option to email the results of the simulation, the email settings can be specific to a particular machine.

Example

{
    "comment__": "Number of threads to use for hyper parameter tuning",
"tuning_threads": 2,
	"comment__": "Number of threads to use for Meta simulation, running a number of simulations and taking the average of them",
"meta_sim_threads": 5,
	"comment__": "Number of simulation samples to compute to use for Meta simulation, running a number of simulations and taking the average of them",
"meta_sim_samples": 5,
	"comment__": "Name of the parameters that we want to tune",
"param_to_tune": "train_forward_dynamics",
	"comment__": "Parameter bounds, for which values will be sampled",
"param_bounds": [1.0, 10.0],
	"comment__": "parameter value type [int|float|bool]",
"param_data_type": "bool",
	"comment__": "Number of samples to evaluation between the given bounds",
"num_param_samples": 2,
	"comment__": "The fully qualified name of the mail server to use",
"mail_server_name": "localhost",
	"comment__": "from email address",
"from_email_address": "csguestn",
	"comment__": "Flag to say this is a testing run",
"testing": false,
	"comment__": "Type of curve to blend values along [linear|squared|exponential]",
"curve_scheme": "linear"
}

Parameters

Name	values	Notes
param_data_type	[int	float
param_bounds	[lower_bound, upper bound]	Specifies the lower and upper bounds when sampling learning parameters. This also interacts with the curve_scheme setting to determine how the blending is performed.
curve_scheme	[linear	squared