US 11,455,576 B2Granted

Architecture for explainable reinforcement learning

AssigneeUMNAI Limited

InventorsAngelo Dalli, Mauro Pirrone, Matthew Grech

Life of this patent

Estimating the term…

Loading patent…

Abstract

An exemplary embodiment may provide an explainable reinforcement learning system. Explanations may be incorporated into an exemplary reinforcement learning agent/model or a corresponding environmental model. The explanations may be incorporated into an agent's state and/or action space. An explainable Bellman equation may implement an explainable state and explainable action as part of an explainable reward function. An explainable reinforcement learning induction method may implement a dataset to provide a white-box model which mimics a black-box reinforcement learning system. An explainable generative adversarial imitation learning model may implement an explainable generative adversarial network to train the occupancy measure of a policy and may generate multiple levels of explanations. Explainable reinforcement learning may be implemented on a quantum computing system using an embodiment of an explainable Bellman equation.

Claims30

Dependent claims

A method for providing an explainable agent for estimating an explainable reward function of a reinforcement learning program and managing a controller according to an estimate of the explainable reward function, comprising:

acquiring, from one or more sensors configured to monitor an external environment, an observed space comprising one or more states and one or more actions, and, with the reinforcement learning program, modeling the observed space as a plurality of explainable state-action pairs;
forming one or more explainable models comprising a simulated environment and a reward function, and integrating at least one of the one or more explainable models with a controller interface associated with the controller;
returning at least one explanation from at least one of the one or more explainable models corresponding to at least one state and at least one reward;
generating an explanation scaffolding;
applying a filter or transform to convert the explanation scaffolding into an interpretation scaffolding;
and
identifying, based on the at least one state, at least one state-action pair, and executing at least one action in the external environment via the controller via the controller interface corresponding to the at least one state-action pair, based on the at least one explanation, wherein the at least one state-action pair is maintained as a constant connection point after selecting the at least one state until after the at least one action is performed.

4 more elements

Dependencies are inferred from the claim text.

Facts

Application: 17/525,395Filed 12 Nov 2021 · us provisional application
Priority: 12 Nov 2020
Published: US 2022/0147876 A112 May 2022
Granted: 27 Sep 2022
Term ends: Not estimatedThe source records don’t support an estimate.

Classification22 codes

G06N 3/042
Architecture, e.g. interconnection topology: knowledge-based neural networks; Logical representations of neural networks
G06N 20/00
Machine learning
G06N 3/0442
Recurrent networks, e.g. Hopfield networks: characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
G06N 3/0455
Combinations of networks: auto-encoder networks; Encoder-decoder networks