Methods and apparatus to generate acceptability criteria for autonomous systems plans

Patent metadata

Field	Value
Publication number	US-2019317455-A1
Application number	US-201916456957-A
Country	US
Kind code	A1
Filing date	Jun 28, 2019
Priority date	Jun 28, 2019
Publication date	Oct 17, 2019
Grant date	—

Abstract

Apparatus, systems, articles of manufacture, and methods to generate acceptability criteria for autonomous systems plans are disclosed. An example apparatus includes a data compiler to compile data generated by the autonomous system into an autonomous system task dataset, a data encoder to encode the dataset for input into a rule distillation neural network architecture, a model trainer to train the rule distillation neural network architecture, an adaptor to adapt the trained rule distillation neural network architecture to a new input data domain using the autonomous system task dataset, a verifier to generate formally verified acceptability criteria, and an inferer to evaluate a control command, the evaluation resulting in an acceptance or rejection of the command.

First claim

1 . An apparatus for validating commands of an autonomous system, the apparatus comprising:
a data compiler to compile data generated by the autonomous system into an autonomous system task dataset; a data encoder to encode the dataset for input into a rule distillation neural network architecture; a model trainer to train the rule distillation neural network architecture; an adaptor to adapt the trained rule distillation neural network architecture to a new input data domain using the autonomous system task dataset; a verifier to generate formally verified acceptability criteria; and an inferer to evaluate a control command, the evaluation resulting in an acceptance or rejection of the command.

2 . The apparatus of claim 1 , wherein the verifier generates formally verifiable criteria for at least one of an embodiment, a task, a sensed state, or a control command.

3 . The apparatus of claim 2 , further including:
a task planner to generate a sequence of control commands, the task planner to receive at least one of an embodiment description or a task description; and a simulator to generate synthetic sensor data from a simulated environment, the simulator to receive the sequence of control commands generated by the task planner.

4 . The apparatus of claim 1 , wherein the adaptor includes a self-supervised adaptation mode to modify the neural network architecture offline.

5 . The apparatus of claim 1 , wherein the new input data domain includes at least one of a new embodiment description or a new task description.

6 . The apparatus of claim 1 , wherein the acceptability criteria are at least one of a safety related or a performance related criteria.

7 . The apparatus of claim 1 , wherein evaluation of a control command is used to determine if the control command is acceptable under at least one of a current state of the system, a current task, or a current environment.

8 . The apparatus of claim 3 , wherein the adaptor is to train the system using the simulator and the task planner.

9 . The apparatus of claim 1 , wherein at least one of the model trainer or adaptor include a cost function, the cost function iteratively optimized during the training, the training completed when the cost function is converged.

10 . The apparatus of claim 1 , wherein the inferer is to evaluate a control command using an on-line inference mode.

11 . The apparatus of claim 1 , wherein the encoder is to encode:
an embodiment, the embodiment input in a mark-up language, the mark-up language encoded into a numerical representation using word embedding, a recurrent neural network to output an encoded embodiment; and a task, the task encoded including at least one of unidimensional sensor values or multidimensional sensor values, wherein the unidimensional sensor values are processed as a sequence of pairs by a recurrent neural network, wherein the multidimensional sensor values are normalized and encoded using the recurrent neural network and a convolutional neural network.

12 . The apparatus of claim 1 , wherein the rule distillation network architecture is to:
concatenate the encoded input, the input fed to a recurrent neural network; output a hidden state, the hidden state converted by a MultiLayer Perceptron (MLP) to a distribution over multiple rule statements; and output, for the rule statement, a sensor identifier, a sensor value, and an instruction, the outputs forming a rule list for formal verification by the verifier.

13 . The apparatus of claim 1 , wherein the verifier includes a temporal logic requirement used to calculate a correctness measure for a reachability analysis used during the neural network training.

14 . A method of validating commands of an autonomous system, the method comprising:
compiling data generated by the autonomous system into an autonomous system task dataset; encoding the dataset for input into a rule distillation neural network architecture; training the rule distillation neural network architecture; modifying the rule distillation neural network architecture by adapting it to a new input data domain, the autonomous system task dataset used to train the modified neural network architecture; generating formally verified acceptability criteria; and evaluating a control command, the evaluation to result in an acceptance or rejection of the command.

15 . The method of claim 14 , wherein the modifying of the rule distillation neural network architecture includes generating a sequence of control commands and synthetic sensor data from a simulated environment.

16 . The method of claim 14 , wherein the formal verification of the acceptability criteria includes a reachability analysis for continuous measure of a robustness of an execution plan to determine constraints towards rules learned by the system.

17 . The method of claim 14 , wherein training the rule distillation neural network includes feeding encoded data inputs into a recurrent neural network, the recurrent neural network generating logic rule statements for use in a rule set.

18 - 21 . (canceled)

22 . The method of claim 14 , wherein the encoding includes:
encoding an embodiment, the embodiment input in a mark-up language, the mark-up language encoded into a numerical representation using word embedding, a recurrent neural network to output an encoded embodiment; and encoding a task, the task encoded including at least one of unidimensional sensor values or multidimensional sensor values, wherein the unidimensional sensor values are processed as a sequence of pairs by a recurrent neural network, wherein the multidimensional sensor values are normalized and encoded using the recurrent neural network and a convolutional neural network.

23 . A non-transitory computer readable storage medium comprising computer readable instructions that, when executed, cause one or more processors to, at least:
compile data generated by an autonomous system into an autonomous system task dataset; encode the dataset for input into a rule distillation neural network architecture; train the rule distillation neural network architecture; modify the rule distillation neural network architecture by adapting it to a new input data domain, the autonomous system task dataset used to train the modified neural network architecture; generate formally verified acceptability criteria; and evaluate a control command, the evaluation resulting in an acceptance or rejection of the command.

24 . The storage medium of claim 23 , wherein the instructions further cause the one or more processors to generate a sequence of control commands and synthetic sensor data from a simulated environment.

25 . The storage medium of claim 23 , wherein the instructions, when executed, cause the one or more processors to feed encoded data inputs into a recurrent neural network, a recurrent neural network generating logic rule statements for use in a rule set.

26 . The storage medium of claim 23 , wherein the instructions, when executed, cause the one or more processors to generate a rule set for each of a command state and a sensed state input data, iterate until a cost function is converged, and train the rule distillation neural network before a new command is executed, the new input data to be provided at every iteration of a control loop.

27 . The storage medium of claim 23 , wherein the instructions, when executed, cause the one or more processors to determin

Assignees

• Intel Corp

Inventors

• Leon Javier Felip
• Turek Javier Sebastian
• Gonzalez Aguirre David I
• Alvarez Ignacio Javier
• Remis Luis Carlos Maria
• Gottschlich Justin

Classifications

• G06N5/025

  Extracting rules from data · CPC title

• G06N3/08

  Learning methods · CPC title

• G06N3/045

  Combinations of networks · CPC title

• G06N3/044

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

• G05B13/027 Primary

  using neural networks only · CPC title