Recursive learning for artificial intelligent agents

Patent metadata

Field	Value
Publication number	US-2020257963-A1
Application number	US-201916557515-A
Country	US
Kind code	A1
Filing date	Aug 30, 2019
Priority date	Feb 13, 2019
Publication date	Aug 13, 2020
Grant date	—

Abstract

The present disclosure relates to a system, and method for computer-based recursive learning of artificial intelligence (AI) apprentice agents. The system includes a system circuitry in communication with a database and a memory. The system circuitry is configured to receive a new data-structure comprising one or more inputs and a goal, and convert, using a perception agent, the one or more inputs of the new data-structure into one or more input feature parameters of the new data-structure. The system circuitry is configured to obtain, using a reasoning agent, an action for the new data-structure, and determine, using an evaluation agent, whether the action for the new data-structure generates the goal of the new data-structure. When it is determined that the action generates the goal of the new data-structure, the system circuitry is further configured to store the new data-structure in the database.

First claim

What is claimed is:

1 . A system for computer-based recursive learning, the system comprising:
a database for storing historically validated data-structures wherein each historical data-structure comprises a goal, a set of sensor data, and a set of corresponding action parameters in response to the set of sensor data for achieving the goal of the historical data-structure; a memory for storing a computer-based recursive e-learning model comprising:
a perception agent,
a reasoning agent, and
an evaluation agent; and
system circuitry in communication with the database and the memory, wherein the system circuitry is configured to:
receive an input data-structure comprising one or more sensor inputs and a goal of the input data-structure,
convert, using the perception agent, the one or more sensor inputs of the input data-structure into an input feature vector of a predetermined dimension,
obtain, using the reasoning agent, a predicted action corresponding to the input feature vector and the goal of the input data-structure based on the historically validated data-structures,
determine, using the evaluation agent, whether the predicted action simulatively produces the goal of the input data-structure, and
when it is determined that the predicted action simulatively produces the goal of the input data-structure, store the input data-structure and the predicted action in the database as part of the historically validated data-structures.

2 . The system according to claim 1 , wherein:
the reasoning agent comprises a retrieval sub-agent and an adaptation sub-agent; and when the system circuitry is configured to obtain, using the reasoning agent, the predicted action, the system circuitry is configured to:
retrieve, using the retrieval sub-agent, an action for the input data-structure corresponding to the input feature vector and the goal of the input data-structure based on the historically validated data-structures, and
adapt, using the adaptation sub-agent, the retrieved action for the input data-structure to obtain an adapted action as the predicted action for the input data-structure.

3 . The system according to claim 2 , wherein, when the system circuitry is configured to retrieve the action for the input data-structure corresponding to the input feature vector and the goal of the input data-structure based on the historically validated data-structures, the system circuitry is configured to:
determine, using the retrieval sub-agent, whether the one or more sensor inputs belong to a non-state-based input category; and when it is determined that the one or more sensor inputs belong to the non-state-based input category,
compute, using the retrieval sub-agent, a similarity between the input feature vector and the goal of the input data-structure and corresponding input feature vector and goals of the historically validated data-structures stored in the database,
obtain, using the retrieval sub-agent, one or more candidate data-structures from the database with their highest values of similarity, and
retrieve, using the retrieval sub-agent, the retrieved action for the input data-structure based on actions of the one or more candidate data-structures from the database.

4 . The system according to claim 3 , wherein:
the retrieval sub-agent computes the similarity between two data-structures based on a K-Nearest Neighbor (KNN) model; and the KNN model comprises a distance metrics configured to compute distance functions, wherein the distance metrics may comprises at least one of Equality, Euclidean Distance, Manhattan distance, Jaccard distance, or Cosine distance.

5 . The system according to claim 2 , wherein, when the system circuitry is configured to retrieve the action for the input data-structure corresponding to the input feature vector and the goal of the input data-structure based on the historically validated data-structures, the system circuitry is configured to:
determine, using the retrieval sub-agent, whether the one or more sensor inputs belong to a non-state-based input category; when it is determined that the one or more sensor inputs belong to the non-state-based input category, retrieve, using the retrieval sub-agent, the action for the input data-structure corresponding to the input feature vector and the goal of the input data-structure based on the historically validated data-structures; and wherein the retrieval sub-agent comprises a supervised algorithm configured to learn a non-linear function for classification based on a deep learning approach.

6 . The system according to claim 5 , wherein:
the retrieval sub-agent comprises a Multilayer Perceptron (MLP) classification model.

7 . The system according to claim 2 , wherein, when the system circuitry is configured to retrieve the action for the input data-structure corresponding to the input feature vector and the goal of the input data-structure based on the historically validated data-structures, the system circuitry is configured to:
determine, using the retrieval sub-agent, whether the one or more sensor inputs belong to a state-based input category; and when it is determined that the one or more sensor inputs belong to the state-based input category, retrieve, by the retrieval sub-agent, the retrieved action for the input data-structure corresponding to the input feature vector and the goal of the input data-structure based on the historically validated data-structures.

8 . The system according to claim 7 , wherein
the retrieval sub-agent retrieves the retrieved action for the input data-structure using a window model; the window model comprises a deep learning based sequential learning approach and a Long Short Term Memory (LSTM) layer to predict the retrieved action; the window model comprises a window size N, N being an integer larger than 1; and the retrieval sub-agent is configured to:
pass the input feature vector as a sequence to the LSTM layer, and
retrieve the retrieved action based on the input data-structure, (N−1) previous actions, and the historically validated data-structures stored in the database.

9 . The system according to claim 7 , wherein
the retrieval sub-agent retrieves the retrieved action for the input data-structure using a human activity recognition model; the human activity recognition model comprises a Recurrent Neural Network (RNN) with bi-directional Long Short-Term Memory cells (LSTMs) to predict the retrieved action; and the retrieval sub-agent is configured to:
pass the input feature vector as a sequence to the bi-directional LSTMs, and
retrieve the retrieved action based on the input data-structure, one or more previous actions, and the historically validated data-structures stored in the database.

10 . The system according to claim 2 , wherein, when the system circuitry is configured to adapt, using the adaptation sub-agent, the retrieved action for the input data-structure to obtain the adapted action as the predicted action for the input data-structure, the system circuitry is configured to:
obtain, using the adaptation sub-agent, action parameters of the retrieved action; and adapt, using the adaptation sub-agent, one or more action parameters of the action parameters to obtain the adapted action.

11 . The system according to claim 2 , wherein,
the adaptation sub-agent comprises at least one of:
a machine learning model comprising at least one of a regression technique or a classification technique;
a rule-based model comprising at least one of an Event-Condition-Action (ECA) semantic form or association rules; or
a recursive-based model configured to sub-divide the input data-structure into one or m

Assignees

• Accenture Global Solutions Ltd

Inventors

• Abhinav Kumar
• Dubey Alpana
• Jain Sakshi
• Arora Veenu
• Sharvani Hindnavis Vijaya

Classifications

• G06V10/776

  Validation; Performance evaluation · CPC title

• G06N3/045

  Combinations of networks · CPC title

• G06N20/00 Primary

  Machine learning · CPC title

• G06N3/044

  Recurrent networks, e.g. Hopfield networks · CPC title

• G06N3/048

  Activation functions · CPC title