Redundant machine learning architecture for high-risk environments
The techniques disclosed herein enable systems to enhance the resilience of autonomous control systems through a fault-tolerant machine learning architecture. To achieve this, a fault-tolerant machine learning agent is constructed with a selector agent, a nominal agent, and a redundancy agent which is a multidimensional lookup table. The fault-tolerant machine learning agent extracts state data from an environment containing a control system and various components. The nominal agent and the redundancy agent generate actions for application to the control system based on the state data which are provided to the selector agent. Based on an analysis of the state data, the selector agent can detect a failure condition. In the event of a failure condition, the selector agent deploys the action generated by the redundancy agent lookup table to resolve the failure condition and restore normal operations.
1. A method comprising:
extracting, by one or more processing units, data defining a current state of an environment containing a plurality of components and a control system that is managed by a fault tolerant machine learning agent comprising a selector agent, a nominal agent, and a redundancy agent;
generating, by the nominal agent, a first action based on the data defining the current state of the environment, wherein the first action:
is generated irrespective of a failure condition of the plurality of components; and
presumes a normal function of the plurality of components;
generating, by the redundancy agent, a second action based on the data defining the current state of the environment, wherein the second action:
is generated irrespective of the failure condition of the plurality of components;
presumes a malfunction of the plurality of components; and
defines an operation for resolving the malfunction;
calculating, by the selector agent, a plurality of error metrics based on the data defining the current state of the environment;
determining, by the selector agent, whether one or more of the plurality of error metrics meets or exceeds a threshold error metric indicating the failure condition of the plurality of components; and
selecting, by the selector agent, the first action generated by the nominal agent for application to the control system or the second action generated by the redundancy agent for application to the control system based on whether the one or more of the plurality of error metrics meets or exceeds the threshold error metric.
2. The method of claim 1 , wherein the redundancy agent is a two-dimensional lookup table.
3. The method of claim 1 , wherein the redundancy agent is constructed based on historical state and action data extracted from the control system.
4. The method of claim 1 , wherein:
the redundancy agent includes a plurality of predetermined actions; and
generating the second action comprises selecting the second action from the plurality of predetermined actions.
5. The method of claim 1 , wherein the second action is generated based on an operating condition of the plurality of components.
6. The method of claim 1 , further comprising constructing the redundancy agent by:
dividing a historical state and action dataset into a plurality of subsets to construct a corresponding plurality of data models;
providing each data model of the plurality of data models to a corresponding machine learning agent to generate a plurality of suggested actions; and
constructing a lookup table comprising the plurality of suggested actions.
7. The method of claim 1 , wherein generating the second action comprises:
providing one or more operating conditions to the redundancy agent; and
indexing a lookup table of the redundancy agent comprising a plurality of predetermined actions with the one or more operating conditions to select the second action from the plurality of predetermined actions.
8. A system comprising:
one or more processing units; and
a computer-readable medium having encoded thereon, computer-readable instructions that when executed by the one or more processing units, cause the system to:
extract, data defining a current state of an environment containing a plurality of components and a control system that is managed by a fault tolerant machine learning agent comprising a selector agent, a nominal agent, and a redundancy agent;
generate, by the nominal agent, a first action based on the data defining the current state of the environment, wherein the first action:
is generated irrespective of a failure condition of the plurality of components; and
presumes a normal function of the plurality of components;
generate, by the redundancy agent, a second action based on the data defining the current state of the environment, wherein the second action:
is generated irrespective of the failure condition of the plurality of components;
presumes a malfunction of the plurality of components; and
defines an operation for resolving the malfunction;
calculate, by the selector agent, a plurality of error metrics based on the data defining the current state of the environment;
determine, by the selector agent, whether one or more of the plurality of error metrics meets or exceeds a threshold error metric indicating the failure condition of the plurality of components; and
select, by the selector agent, the first action generated by the nominal agent for application to the control system or the second action generated by the redundancy agent for application to the control system based on whether the one or more of the plurality of error metrics meets or exceeds the threshold error metric.
9. The system of claim 8 , wherein the redundancy agent is a two-dimensional lookup table.
10. The system of claim 8 , wherein the redundancy agent is constructed based on historical state and action data extracted from the control system.
11. The system of claim 8 , wherein:
the redundancy agent includes a plurality of predetermined actions; and
generating the second action comprises selecting the second action from the plurality of predetermined actions.
12. The system of claim 8 , wherein the second action is generated based on an operating condition of the plurality of components.
13. The system of claim 8 , wherein the computer-readable instructions further cause the system to construct the redundancy agent by:
dividing a historical state and action dataset into a plurality of subsets to construct a corresponding plurality of data models;
providing each data model of the plurality of data models to a corresponding machine learning agent to generate a plurality of suggested actions; and
constructing a lookup table comprising the plurality of suggested actions.
14. The system of claim 8 , wherein generating the second action comprises:
providing one or more operating conditions to the redundancy agent; and
indexing a lookup table of the redundancy agent comprising a plurality of predetermined actions with the one or more operating conditions to select the second action from the plurality of predetermined actions.
15. A computer-readable storage medium having encoded thereon computer-readable instructions that when executed by one or more processing units cause a system to:
extract, data defining a current state of an environment containing a plurality of components and a control system that is managed by a fault tolerant machine learning agent comprising a selector agent, a nominal agent, and a redundancy agent;
generate, by the nominal agent, a first action based on the data defining the current state of the environment, wherein the first action:
is generated irrespective of a failure condition of the plurality of components; and
presumes a normal function of the plurality of components;
generate, by the redundancy agent, a second action based on the data defining the current state of the environment, wherein the second action:
is generated irrespective of the failure condition of the plurality of components;
presumes a malfunction of the plurality of components; and
defines an operation for resolving the malfunction;
calculate, by the selector agent, a plurality of error metrics based on the data defining the current state of the environment;
determine, by the selector agent, whether one or more of the plurality of error metrics meets or exceeds a threshold error metric indicating the failure condition of the plurality of components; and
select, by the selector agent, the first action generated by the nominal agent for application to the control system or the second action generated by the redundancy agent for application to the control system based on whether the one or more of the plurality of error metrics meets or exceeds the threshold error metric.
16. The computer-readable storage medium of claim 15 , wherein the redundancy agent is a two-dimensional lookup table.
17. The computer-readable storage medium of claim 15 , wherein the redundancy agent is constructed based on historical state and action data extracted from the control system.
18. The computer-readable storage medium of claim 15 , wherein:
the redundancy agent includes a plurality of predetermined actions; and
generating the second action comprises selecting the second action from the plurality of predetermined actions.
19. The computer-readable storage medium of claim 15 , wherein the second action is generated based on an operating condition of the plurality of components.
20. The computer-readable storage medium of claim 15 , wherein the computer-readable instructions further cause the system to construct the redundancy agent by:
dividing a historical state and action dataset into a plurality of subsets to construct a corresponding plurality of data models;
providing each data model of the plurality of data models to a corresponding machine learning agent to generate a plurality of suggested actions; and
constructing a lookup table comprising the plurality of suggested actions.