Systems and methods for machine learning using a trusted model

A machine learning system for creating a trusted model that improves operation of a computer system controller, comprising:

• a user interface configured to receive at least one trust-related constraint;
• a model training module configured to train a model using input data and machine learning algorithms;
• a model checking module configured to determine whether the model satisfies the at least one trust-related constraint;
• and
• a repair module, wherein the repair module includes at least one of:
• (a) a model repair module that modifies the model using one or more model repair algorithms;
• (b) a data repair module that modifies the input data using one or more data repair algorithms;
• or
• (c) a reward repair module that modifies a reward function of the model using one or more reward repair algorithms,
• wherein the machine learning system is configured to provide the trained model to the computer system controller as a trusted model that improves the operation of the computer system controller by enabling the computer system controller to perform the at least one system action within predetermined guarantees when it is determined that the model satisfies the at least one trust-related constraint.

A non-transitory computer readable medium for storing computer instructions that, when executed by at least one processor causes the at least one processor to perform a method of using machine learning to create a trusted model that improves operation of a computer system controller, the method comprising:

• (a) determining at least one trust-related constraint related to at least one system action that could be performed by the computer system controller;
• (b) training a model of the computer system controller using an initial set of input data;
• (c) extracting the model;
• (d) determining whether the model satisfies the at least one trust-related constraint;
• (e) when it is determined that the model does not satisfy the at least one trust-related constraint:
• (1) modifying at least one of:
• the model using one or more model repair algorithms,
• the input data using one or more data repair algorithms, or
• a reward function of the model using one or more reward repair algorithms;
• and
• (2) re-training the model using at least one of the modified model, the modified input data, or the modified reward function;
• and
• (f) when it is determined that the model satisfies the at least one trust-related constraint, providing the model as a trusted model that improves the operation of the computer system controller by enabling the computer system controller to perform the at least one system action within predetermined guarantees.

A method for using machine learning to create a trusted model that improves operation of an autonomous vehicle controller, the method comprising:

• receiving a temporal driving constraint that must be satisfied before the autonomous vehicle controller can perform an action;
• training a probabilistic model of the autonomous vehicle controller using an initial set of input data;
• determining whether the probabilistic model satisfies the temporal driving constraint;
• when it is determined that the probabilistic model satisfies the temporal driving constraint, providing the model for use by the autonomous vehicle controller;
• when it is determined that the probabilistic model does not satisfy the temporal driving constraint:
• (1) modifying at least one of:
• the model using one or more model repair algorithms,
• the input data using one or more data repair algorithms, or
• a reward function of the model using one or more reward repair algorithms;
• and
• (2) re-training the model using at least one of the modified model, the modified input data, or the modified reward function;
• and
• (3) once it has been determined that the re-trained model satisfies the temporal driving constraint, providing the model to the autonomous vehicle controller as a trusted model that improves the operation of the autonomous vehicle controller by enabling the autonomous vehicle controller to perform the at least one system action within predetermined guarantees.