Artificial intelligence based system and method for dynamic goal planning

Patent metadata

Field	Value
Publication number	US-2021256434-A1
Application number	US-202016794786-A
Country	US
Kind code	A1
Filing date	Feb 19, 2020
Priority date	Feb 19, 2020
Publication date	Aug 19, 2021
Grant date	—

Abstract

The disclosed system and method provide a way to create, update, and execute dynamic goal plans. Updating a dynamic goal plan may be based on the initial sequence of actions of the goal plan as well as the corresponding states of the actions. By using a sequence to sequence model, a goal plan can still be processed when the length of the input (initial sequence of actions) differs from the length of the output (updated sequence of actions). A sequence to sequence model can determine the interdependencies between actions that can contribute to the optimal order in which actions can efficiently be performed. A single layer neural network or clustering can be used to approximate the state of a goal plan that may be capable infinite states. This approximation improves accuracy in capturing the state of a goal plan, thereby improving accuracy in predicting the future state of a system, which can help with planning (e.g., gathering resources in advance). Projects involving collaboration between virtual and/or human assistants can greatly benefit from the ability to update a dynamic goal plan in real time.

First claim

We claim:

1 . A computer implemented method of updating a dynamic goal plan, comprising:
receiving an initial goal plan comprising an initial action sequence including a plurality of actions ordered in a forward direction; processing the initial action sequence through an encoder of a bidirectional recurrent neural network (“RNN”) to generate an encoder output, including a first hidden state representation; processing the encoder output through a decoder of the bidirectional RNN to generate a decoder output, including a forward hidden state representation and a backward hidden state representation for each action of the initial action sequence; applying a context vector to the decoder output to generate a weighted decoder output; obtaining a state of the initial goal plan, wherein the state of the initial goal plan includes a plurality of states each corresponding to an action of the initial goal plan; converting the state of the initial goal plan into vector embeddings; concatenating the weighted decoder output with the vector embeddings; and processing the concatenated weighted decoder output and vector embeddings through a SoftMax classifier to determine an updated goal plan.

2 . The computer implemented method of claim 1 , further comprising creating, by a goal plan module, the initial goal plan.

3 . The computer implemented method of claim 1 , wherein the bidirectional RNN includes one of a long short-term memory (LSTM) and a gated recurrent unit (GRU).

4 . The computer implemented method of claim 1 , wherein converting the state of the initial goal plan into vector embeddings includes learning vector embeddings through a single layer neural network.

5 . The computer implemented method of claim 1 , wherein converting the state of the initial goal plan into vector embeddings comprises:
clustering a plurality of known states for the initial goal plan and labeling the clusters.

6 . The computer implemented method of claim 1 , wherein the initial goal plan includes a different number of actions from the updated goal plan.

7 . The computer implemented method of claim 1 , wherein the order of the actions in the initial goal plan differs from the order of the actions in the updated goal plan.

8 . A non-transitory computer-readable medium storing software comprising instructions executable by one or more computers which, upon such execution, cause the one or more computers to:
receive an initial goal plan comprising an initial action sequence including a plurality of actions ordered in a forward direction; process the initial action sequence through an encoder of a bidirectional recurrent neural network (“RNN”) to generate an encoder output, including a first hidden state representation; process the encoder output through a decoder of the bidirectional RNN to generate a decoder output, including a forward hidden state representation and a backward hidden state representation for each action of the initial action sequence; apply a context vector to the decoder output to generate a weighted decoder output; obtain a state of the initial goal plan, wherein the state of the initial goal plan includes a plurality of states each corresponding to an action of the initial goal plan; convert the state of the initial goal plan into vector embeddings; concatenate the weighted decoder output with the vector embeddings; and process the concatenated weighted decoder output and vector embeddings through a SoftMax classifier to determine an updated goal plan.

9 . The non-transitory computer-readable medium storing software of claim 8 , wherein the instructions further cause the one or more computers to create, by a goal plan module, the initial goal plan.

10 . The non-transitory computer-readable medium storing software of claim 8 , wherein the bidirectional RNN includes one of a long short-term memory (LSTM) and a gated recurrent unit (GRU).

11 . The non-transitory computer-readable medium storing software of claim 8 , wherein converting the state of the initial goal plan into vector embeddings includes learning vector embeddings through a single layer neural network.

12 . The non-transitory computer-readable medium storing software of claim 8 , wherein converting the state of the initial goal plan into vector embeddings comprises:
clustering a plurality of known states for the initial goal plan and labeling the clusters.

13 . The non-transitory computer-readable medium storing software of claim 8 , wherein the initial goal plan includes a different number of actions from the updated goal plan.

14 . The non-transitory computer-readable medium storing software of claim 8 , wherein the order of the actions in the initial goal plan differs from the order of the actions in the updated goal plan.

15 . A system for updating a dynamic goal plan, comprising:
one or more computers and one or more storage devices storing instructions that are operable, when executed by the one or more computers, to cause the one or more computers to:
receive an initial goal plan comprising an initial action sequence including a plurality of actions ordered in a forward direction;
process the initial action sequence through an encoder of a bidirectional recurrent neural network (“RNN”) to generate an encoder output, including a first hidden state representation;
process the encoder output through a decoder of the bidirectional RNN to generate a decoder output, including a forward hidden state representation and a backward hidden state representation for each action of the initial action sequence;
apply a context vector to the decoder output to generate a weighted decoder output;
obtain a state of the initial goal plan, wherein the state of the initial goal plan includes a plurality of states each corresponding to an action of the initial goal plan;
convert the state of the initial goal plan into vector embeddings;
concatenate the weighted decoder output with the vector embeddings; and
process the concatenated weighted decoder output and vector embeddings through a SoftMax classifier to determine an updated goal plan.

16 . The system of claim 15 , wherein the instructions further cause the one or more computers to create, by a goal plan module, the initial goal plan.

17 . The system of claim 15 , wherein the bidirectional RNN includes one of a long short-term memory (LSTM) and a gated recurrent unit (GRU).

18 . The system of claim 15 , wherein converting the state of the initial goal plan into vector embeddings includes learning vector embeddings through a single layer neural network.

19 . The system of claim 15 , wherein converting the state of the initial goal plan into vector embeddings comprises:
clustering a plurality of known states for the initial goal plan and labeling the clusters.

20 . The system of claim 15 , wherein the initial goal plan includes a different number of actions from the updated goal plan.

Assignees

• Accenture Global Solutions Ltd

Inventors

• Dubey Alpana
• Abhinav Kumar
• Jain Sakshi
• Arora Veenu

Classifications

• G06N3/044

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

• G06N3/045

  Combinations of networks · CPC title

• G06N3/0442

  characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU] · CPC title

• G06N3/0455

  Auto-encoder networks; Encoder-decoder networks · CPC title

• G06N3/09

  Supervised learning · CPC title