Method and apparatus for constructing informative outcomes to guide multi-policy decision making
In Multi-Policy Decision-Making (MPDM), many computationally-expensive forward simulations are performed in order to predict the performance of a set of candidate policies. In risk-aware formulations of MPDM, only the worst outcomes affect the decision making process, and efficiently finding these influential outcomes becomes the core challenge. Recently, stochastic gradient optimization algorithms, using a heuristic function, were shown to be significantly superior to random sampling. In this disclosure, it was shown that accurate gradients can be computed-even through a complex forward simulation—using approaches similar to those in dep networks. The proposed approach finds influential outcomes more reliably, and is faster than earlier methods, allowing one to evaluate more policies while simultaneously eliminating the need to design an easily-differentiable heuristic function.
1. A method, comprising:
operating a vehicle according to a first policy;
while operating the vehicle according to the first policy, evaluating a set of policy options, comprising:
detecting a set of objects in the vehicle's environment;
evaluating each of the set of policy options, comprising, for each of the set of policy options:
identifying a set of multiple potential outcomes associated with the policy, comprising guiding the set of multiple potential outcomes to a particular outcome category;
evaluating the set of multiple potential outcomes to produce a score;
selecting a second policy from the set of policy options based on a set of scores comprising the produced score for each policy option; and
operating the vehicle according to the second policy;
wherein evaluating the set of multiple potential outcomes comprises, for each of the set of policy options, predicting a progress of the vehicle toward a predetermined goal and each of the set of policy options is further evaluated based on a quantified risk of executing the policy option.
2. The method of claim 1 , wherein the particular outcome category comprises a category associated with a high risk of collision between one or more of: the vehicle and at least one object of the set of objects; or a first object and a second object of the set of objects.
3. The method of claim 2 , wherein guiding the set of multiple potential outcomes to the particular outcome category comprises adjusting input data associated with the set of objects.
4. The method of claim 3 , wherein the input data comprises at least one of position or motion information associated with the set of objects.
5. The method of claim 3 , wherein the input data comprises a goal associated with the set of objects.
6. The method of claim 3 , wherein guiding the set of multiple potential outcomes further comprises applying a backpropagation process.
7. The method of claim 1 , wherein evaluating the set of multiple potential outcomes comprises, for each of the set of policy options, predicting a quantified inconvenience to the set of objects in response to executing the policy option, and producing the score based on the quantified inconvenience.
8. The method of claim 7 , wherein the quantified inconvenience is calculated based on a set of predicted distances between the vehicle and a closest object of the set of objects.
9. The method of claim 7 , wherein each of the set of policy options is further evaluated based on a quantified risk of executing the policy option.
10. A system, comprising:
a processing subsystem of a controlled vehicle configured to:
evaluate a set of policy options, comprising:
detecting a set of objects in the controlled vehicle's environment;
evaluating each of the set of policy options, comprising, for each of the set of policy options:
identifying a set of multiple potential outcomes associated with the policy, comprising guiding the set of multiple potential outcomes to a particular outcome category;
evaluating the set of multiple potential outcomes to produce a score;
selecting a policy from the set of policy options based on a set of scores comprising the produced score for each policy option; and
a control subsystem configured to implement the selected policy;
wherein evaluating the set of multiple potential outcomes comprises, for each of the set of policy options, predicting a quantified inconvenience to the set of objects in response to executing the policy option, and producing the score based on the quantified inconvenience.
11. The system of claim 10 , wherein the particular outcome category comprises a category associated with a high risk of collision between one or more of: the controlled vehicle and at least one object of the set of objects; or a first object and a second object of the set of objects.
12. The system of claim 11 , wherein guiding the set of multiple potential outcomes to the particular outcome category comprises adjusting input data associated with the set of objects.
13. The system of claim 12 , wherein the input data comprises at least one of position or motion information associated with the set of objects.
14. The system of claim 12 , wherein the input data comprises a goal associated with the set of objects.
15. The system of claim 10 , wherein the quantified inconvenience is calculated based on a set of predicted distances between the controlled vehicle and a closest object of the set of objects.
16. The system of claim 10 , wherein each of the set of policy options is further evaluated based on a quantified risk of executing the policy option.
17. The system of claim 10 , wherein evaluating the set of multiple potential outcomes comprises, for each of the set of policy options, predicting a progress of the vehicle toward a predetermined goal.
18. The system of claim 10 , wherein guiding the set of multiple potential outcomes includes applying a backpropagation process.
19. The system of claim 10 wherein evaluating the set of multiple potential outcomes comprises, for each of the set of policy options, predicting a progress of the vehicle toward a predetermined goal and each of the set of policy options is further evaluated based on a quantified risk of executing the policy option.
20. A method, comprising:
operating a vehicle according to a first policy;
while operating the vehicle according to the first policy, evaluating a set of policy options, comprising:
detecting a set of objects in the vehicle's environment;
evaluating each of the set of policy options, comprising, for each of the set of policy options:
identifying a set of multiple potential outcomes associated with the policy, comprising guiding the set of multiple potential outcomes to a particular outcome category;
evaluating the set of multiple potential outcomes to produce a score;
selecting a second policy from the set of policy options based on a set of scores comprising the produced score for each policy option; and
operating the vehicle according to the second policy;
wherein the particular outcome category comprises a category associated with a high risk of collision between one or more of: the vehicle and at least one object of the set of objects; or a first object and a second object of the set of objects; and;
wherein guiding the set of multiple potential outcomes to the particular outcome category comprises adjusting input data associated with the set of objects and applying a backpropagation process.
21. A method, comprising:
operating a vehicle according to a first policy;
while operating the vehicle according to the first policy, evaluating a set of policy options, comprising:
detecting a set of objects in the vehicle's environment;
evaluating each of the set of policy options, comprising, for each of the set of policy options:
identifying a set of multiple potential outcomes associated with the policy, comprising guiding the set of multiple potential outcomes to a particular outcome category;
evaluating the set of multiple potential outcomes to produce a score;
selecting a second policy from the set of policy options based on a set of scores comprising the produced score for each policy option; and
operating the vehicle according to the second policy;
wherein evaluating the set of multiple potential outcomes comprises, for each of the set of policy options, predicting a quantified inconvenience to the set of objects in response to executing the policy option, and producing the score based on the quantified inconvenience.
22. A system, comprising:
a processing subsystem of a controlled vehicle configured to:
evaluate a set of policy options, comprising:
detecting a set of objects in the controlled vehicle's environment;
evaluating each of the set of policy options, comprising, for each of the set of policy options:
identifying a set of multiple potential outcomes associated with the policy, comprising guiding the set of multiple potential outcomes to a particular outcome category;
evaluating the set of multiple potential outcomes to produce a score;
selecting a policy from the set of policy options based on a set of scores comprising the produced score for each policy option; and
a control subsystem configured to implement the selected policy;
wherein the particular outcome category comprises a category associated with a high risk of collision between one or more of: the vehicle and at least one object of the set of objects; or a first object and a second object of the set of objects; and
wherein guiding the set of multiple potential outcomes to the particular outcome category comprises adjusting input data associated with the set of objects and applying a backpropagation process.
23. A system, comprising:
a processing subsystem of a controlled vehicle configured to:
evaluate a set of policy options, comprising:
detecting a set of objects in the controlled vehicle's environment;
evaluating each of the set of policy options, comprising, for each of the set of policy options:
identifying a set of multiple potential outcomes associated with the policy, comprising guiding the set of multiple potential outcomes to a particular outcome category;
evaluating the set of multiple potential outcomes to produce a score;
selecting a policy from the set of policy options based on a set of scores comprising the produced score for each policy option; and
a control subsystem configured to implement the selected policy;
wherein evaluating the set of multiple potential outcomes comprises, for each of the set of policy options, predicting a progress of the vehicle toward a predetermined goal and each of the set of policy options is further evaluated based on a quantified risk of executing the policy option.