Multi-objective bayesian optimization of machine learning model for autonomous vehicle operation
Machine learning model optimization systems and methods are disclosed. A system receives sensor data captured by one or more sensors of a vehicle during a first time period. The vehicle uses a first trained machine learning (ML) model for one or more decisions of a first decision type during the first time period. The system generates a second trained ML model at least in part by using the sensor data to train the second trained ML model. The system identifies an optimal trained ML model from a plurality of trained ML models. The plurality of trained ML models includes the first trained ML model and the second trained ML model. The system causes the vehicle to use the optimal trained ML model for one or more further decisions of the first decision type during a second time period after the first time period.
1. A system for machine learning model optimization, the system comprising:
one or more memory units storing instructions; and
one or more processors coupled to a vehicle, wherein execution of the instructions by the one or more processors causes the one or more processors to:
receive sensor data captured by one or more sensors of the vehicle during a first time period, wherein the vehicle uses a first trained machine learning (ML) model to make one or more decisions of a decision type during the first time period;
generate a one-shot trained ML model at least in part by using the sensor data to train the one-shot trained ML model;
generate a plurality of candidate deep learning (DL) architectures using the one-shot trained ML model;
generate an optimal trained ML model based on at least one selected candidate DL architecture of the plurality of candidate DL architectures; and
cause the vehicle to use the optimal trained ML model to make one or more further decisions of the decision type during a second time period after the first time period.
2. The system of claim 1 , wherein the plurality of candidate DL architectures form a Pareto front, wherein generating the optimal trained ML model based on the at least one selected candidate DL architecture includes selecting the at least one selected candidate DL architecture from the plurality of candidate DL architectures based on Pareto optimality corresponding to the Pareto front.
3. The system of claim 1 , wherein the one or more sensors of the vehicle include an image sensor of a camera, and wherein the sensor data includes image data captured by the image sensor.
4. The system of claim 1 , wherein the one or more sensors of the vehicle include a depth sensor, and wherein the sensor data includes depth data captured by the depth sensor.
5. The system of claim 1 , wherein the one or more sensors of the vehicle include a pose sensor, and wherein the sensor data includes pose data captured by the pose sensor.
6. The system of claim 1 , wherein the decision type is associated with object recognition of an object other than the vehicle, wherein the vehicle and the object are both in an environment.
7. The system of claim 1 , wherein the decision type is associated with at least one of tracking movement or predicting of an object other than the vehicle, wherein the vehicle and the object are both in an environment.
8. The system of claim 1 , wherein the decision type is associated with mapping an environment that the vehicle is positioned within, wherein the environment includes the vehicle and one or more other objects.
9. The system of claim 1 , wherein the decision type is associated with generating a route for the vehicle to traverse through an environment that the vehicle is positioned within, wherein the environment includes the vehicle and one or more other objects.
10. The system of claim 9 , wherein execution of the instructions by the one or more processors causes the one or more processors to:
cause the vehicle to actuate one or more actuators of the vehicle to autonomously traverse the route.
11. The system of claim 1 , wherein the decision type is associated with control over one or more actuators that control movement of the vehicle throughout an environment.
12. The system of claim 1 , wherein, to generate the optimal trained ML model based on the at least one selected candidate DL architecture, the one or more processors are configured to train the optimal trained ML model from scratch.
13. The system of claim 1 , wherein, to generate the optimal trained ML model based on the at least one selected candidate DL architecture, the one or more processors are configured to generate the optimal trained ML model based also on one or more parameters of the vehicle.
14. The system of claim 1 , wherein causing the vehicle to use the optimal trained ML model to make the one or more further decisions of the decision type during the second time period includes transmitting the optimal trained ML model to the vehicle.
15. The system of claim 1 , wherein execution of the instructions by the one or more processors causes the one or more processors to:
cause the vehicle to use the optimal trained ML model to make a decision of a second decision type during the second time period, wherein the second decision type is distinct from the decision type.
16. The system of claim 1 , wherein execution of the instructions by the one or more processors causes the one or more processors to:
determine that the optimal trained ML model is more optimal than the first trained ML model for the decision type, wherein the one or more processors are configured to cause the vehicle to use the optimal trained ML model to make the one or more further decisions of the decision type during the second time period based on the optimal trained ML model being more optimal than the first trained ML model for the decision type.
17. The system of claim 1 , wherein the optimal trained ML model is more accurate than the first trained ML model for the decision type.
18. The system of claim 1 , wherein the optimal trained ML model is faster than the first trained ML model for the decision type.
19. A method for machine learning model optimization, the method comprising:
receiving sensor data captured by one or more sensors of a vehicle during a first time period, wherein the vehicle uses a first trained machine learning (ML) model to make one or more decisions of a first decision type during the first time period;
generating a one-shot trained ML model at least in part by using the sensor data to train the one-shot trained ML model;
generating a plurality of candidate deep learning (DL) architectures using the one-shot trained ML model;
generating an optimal trained ML model based on at least one selected candidate DL architecture of the plurality of candidate DL architectures; and
causing the vehicle to use the optimal trained ML model to make one or more further decisions of the first decision type during a second time period after the first time period.
20. A non-transitory computer readable storage medium having embodied thereon a program, wherein the program is executable by a processor to perform a method of machine learning model optimization, the method comprising:
receiving sensor data captured by one or more sensors of a vehicle during a first time period, wherein the vehicle uses a first trained machine learning (ML) model for one or more decisions of a first decision type during the first time period;
generating a one-shot trained ML model at least in part by using the sensor data to train the one-shot trained ML model;
generating a plurality of candidate deep learning (DL) architectures using the one-shot trained ML model;
generating an optimal trained ML model based on at least one selected candidate DL architecture of the plurality of candidate DL architectures; and
causing the vehicle to use the optimal trained ML model to make one or more further decisions of the first decision type during a second time period after the first time period.