Iterative model training and deployment for automated learning systems

1. A computer-implemented method comprising:

receiving a request at an endpoint of a reinforcement learning service to perform inference using a first version of a model on data received from a client application;

returning inference data based on the inference;

logging the inference data in a first data store;

generating training data based on the inference data and feedback data received from the client application using an identifier associated with the inference data and the feedback data;

generating a second version of the model using the training data; and

deploying the second version of the model to the endpoint of the reinforcement learning service by:

loading the second version of the model to the endpoint of the reinforcement learning service;

while the second version of model is being loaded, receiving one or more inference requests at the endpoint of the reinforcement learning service;

queuing the one or more inference requests that are received while the second version of model is being loaded; and

after the second version of model is ready to receive requests, processing the one or more inference requests using the second version of the model.

2. The computer-implemented method of claim 1 , further comprising:

iteratively training further versions of the model based on feedback from a previous version of the model received from the client; and

deploying each further version of the model to the endpoint of the reinforcement learning service.

3. The computer-implemented method of claim 1 , wherein generating training data based on the inference data and feedback data received from the client application using an identifier associated with the inference data and the feedback data further comprises joining a first dataset that includes the inference data and the identifier with a second dataset that includes the feedback data and the identifier, wherein the first dataset and the second dataset each include data from a time period specified by the client application.

4. A computer-implemented method comprising:

generating training data based on inference data, provided by a first version of a model hosted at an endpoint of a machine learning service, and feedback data, received from a client application, using an identifier associated with the inference data and the feedback data;

generating a second version of the model using the training data; and

deploying the second version of the model to the endpoint of the machine learning service by:

loading the second version of the model to the endpoint of the machine learning service;

while the second version of model is being loaded, receiving one or more inference requests at the endpoint of the machine learning service;

queuing the one or more inference requests that are received while the second version of model is being loaded; and

after the second version of model is ready to receive requests, processing the one or more inference requests using the second version of the model.

5. The computer-implemented method of claim 4 , wherein the first version of the model is trained based at least on client training data.

6. The computer-implemented method of claim 5 , wherein the client training data includes an inference, an inference probability, and state information.

7. The computer-implemented method of claim 4 , wherein generating training data based on inference data, provided by a first version of a model hosted at an endpoint of a machine learning service, and feedback data, received from a client application, using an identifier associated with the inference data and the feedback data, further comprises:

joining a first dataset that includes the inference data and the identifier with a second dataset that includes the feedback data and the identifier, wherein the first dataset and the second dataset each include data from a time period specified by the client application.

8. The computer-implemented method of claim 7 , wherein the time period specified by the client application is a time period subsequent to a previous training of the model.

9. The computer-implemented method of claim 4 , wherein the identifier is a universally unique identifier (UUID), wherein each inference in the inference data is assigned its own UUID by the machine learning service and wherein the client application applies the UUID to each corresponding feedback from the feedback data.

10. The computer-implemented method of claim 4 , wherein the training data is sampled from a plurality of different collection time periods.

11. The computer-implemented method of claim 4 , further comprising:

determining the second version of the model has been added to a storage location accessible to the machine learning service; and

loading the second version of the model into memory at the endpoint of the machine learning service.

12. The computer-implemented method of claim 4 , wherein the model returns a random result in response to a subset of received inference requests and an inference result in response to a remainder of received inference requests.

13. A system comprising:

a machine learning service implemented by a first one or more electronic devices; and

a reinforcement learning service implemented by a second one or more electronic devices, the reinforcement learning service including instructions that upon execution cause the reinforcement learning service to:

generate training data based on inference data, provided by a first version of a model hosted at an endpoint of the machine learning service, and feedback data, received from a client application, using an identifier associated with the inference data and the feedback data;

generate a second version of the model using the training data; and

deploy the second version of the model to the endpoint of the machine learning service by:

loading the second version of the model to the endpoint of the machine learning service;

while the second version of model is being loaded, receiving one or more inference requests at the endpoint of the machine learning service;

queuing the one or more inference requests that are received while the second version of model is being loaded; and

after the second version of model is ready to receive requests, processing the one or more inference requests using the second version of the model.

14. The system of claim 13 , wherein the first version of the model is trained based at least on client training data.

15. The system of claim 14 , wherein the client training data includes an inference, an inference probability, and state information.

16. The system of claim 13 , wherein generating training data based on inference data, provided by a first version of a model hosted at an endpoint of the machine learning service, and feedback data, received from a client application, using an identifier associated with the inference data and the feedback data, further comprises:

joining a first dataset that includes the inference data and the identifier with a second dataset that includes the feedback data and the identifier, wherein the first dataset and the second dataset each include data from a time period specified by the client application.

17. The system of claim 16 , wherein the time period specified by the client application is a time period subsequent to a previous training of the model.

18. The system of claim 13 , wherein the identifier is a universally unique identifier (UUID), wherein each inference in the inference data is assigned its own UUID by the machine learning service and wherein the client application applies the UUID to each corresponding feedback from the feedback data.

19. The system of claim 13 , wherein the training data is sampled from a plurality of different collection time periods.

20. The system of claim 13 , further comprising:

determining the second version of the model has been added to a storage location accessible to the machine learning service; and

loading the second version of the model into memory at the endpoint.