Determining action selection policies of an execution device

1. A computer-implemented method of an execution device for generating an action selection policy for completing a task in an environment that includes the execution device and one or more other devices, the method comprising:

at each of a plurality of iterations and for each action among a plurality of possible actions in a state of the execution device in a current iteration, wherein the state of the execution device results from a history of actions taken by the execution device towards completing the task,

identifying an iterative action selection policy of the action in the state of the execution device in a previous iteration, wherein the iterative action selection policy of the action in the state of the execution device in the previous iteration represents a probability of the action taken by the execution device in the state in the previous iteration;

computing a regret value of the action in the state of the execution device in the previous iteration based on the iterative action selection policy of the action in the state of the execution device in the previous iteration, wherein the regret value of the action in the state of the execution device represents a difference between a gain of the execution device attributed to the action towards completing the task after taking the action in the state and a gain of the execution device in the state attributed to the state towards completing the task;

computing an incremental action selection policy of the action in the state of the execution device in the current iteration based on the regret value of the action in the state of the execution device in the previous iteration but not any regret value of the action in the state of the execution device in any iteration prior to the previous iteration;

computing an iterative action selection policy of the action in the state of the execution device in the current iteration based on a weighted sum of the iterative action selection policy of the action in the state of the execution device in the previous iteration and the incremental action selection policy of the action in the state of the execution device in the current iteration;

storing the iterative action selection policy of the action in the state of the execution device in the current iteration; and

in response to determining that a convergence condition is met, outputting the iterative action selection policy of the action in the state of the execution device in the current iteration such that operations of the execution device are controlled according to the iterative action selection policy of the action in the state of the execution device in the current iteration for completing the task.

2. The method of claim 1 , further comprising, in response to determining that the convergence condition is met, controlling the operations of the execution device according to the iterative action selection policy of the action in the state of the execution device in the current iteration for completing the task.

3. The method of claim 1 , wherein computing a regret value of the action in the state of the execution device in the previous iteration based on the iterative action selection policy of the action in the state of the execution device in the previous iteration comprises computing the regret value of the action in the state of the execution device in the previous iteration based on the iterative action selection policy of the action in the state of the execution device in the previous iteration but not any regret value of the action in the state of the execution device in any iteration prior to the previous iteration.

4. The method of claim 1 , wherein computing a regret value of the action in the state of the execution device in the previous iteration based on the iterative action selection policy of the action in the state of the execution device in the previous iteration comprises computing the regret value of the action in the state of the execution device in the previous iteration based on a difference between a counterfactual value of the action in the state of the execution device and a counterfactual value of the state of the execution device, wherein the counterfactual value of the action in the state of the execution device and the counterfactual value of the state of the execution device are computed by recursively traversing a game tree that represents the environment that includes the execution device and the one or more other devices in the previous iteration.

5. The method of claim 1 , wherein the weighted sum of the iterative action selection policy of the action in the state of the execution device in the previous iteration and the incremental action selection policy of the action in the state of the execution device in the current iteration comprises a sum of:

the iterative action selection policy of the action in the state of the execution device in the previous iteration scaled by a first learning rate in the current iteration; and

the incremental action selection policy of the action in the state of the execution device in the current iteration scaled by a second learning rate in the current iteration, and

wherein the first learning rate approaches 1 as the number of iterations, t, approaches infinity, and the second learning rate approaches 0 as t approaches infinity.

6. The method of claim 5 , wherein the first learning rate is (t−1)/t, and the second learning rate is 1/t.

7. The method of claim 1 , further comprising initializing the iterative action selection policy of the action in the state of the execution device in the first iteration based on a uniform random action selection policy.

8. A system for generating an action selection policy for completing a task in an environment that includes an execution device and one or more other devices, the system comprising:

one or more processors; and

one or more computer-readable memories coupled to the one or more processors and having instructions stored thereon that are executable by the one or more processors to perform operations comprising:

at each of a plurality of iterations and for each action among a plurality of possible actions in a state of the execution device in a current iteration, wherein the state of the execution device results from a history of actions taken by the execution device,

identifying an iterative action selection policy of the action in the state of the execution device in a previous iteration, wherein the iterative action selection policy of the action in the state of the execution device in the previous iteration represents a probability of the action taken by the execution device in the state in the previous iteration;

computing a regret value of the action in the state of the execution device in the previous iteration based on the iterative action selection policy of the action in the state of the execution device in the previous iteration, wherein the regret value of the action in the state of the execution device represents a difference between a gain of the execution device attributed to the action towards completing the task after taking the action in the state and a gain of the execution device in the state attributed to the state towards completing the task;

computing an incremental action selection policy of the action in the state of the execution device in the current iteration based on the regret value of the action in the state of the execution device in the previous iteration but not any regret value of the action in the state of the execution device in any iteration prior to the previous iteration;

computing an iterative action selection policy of the action in the state of the execution device in the current iteration based on a weighted sum of the iterative action selection policy of the action in the state of the execution device in the previous iteration and the incremental action selection policy of the action in the state of the execution device in the current iteration;

storing the iterative action selection policy of the action in the state of the execution device in the current iteration; and

in response to determining that a convergence condition is met, outputting the iterative action selection policy of the action in the state of the execution device in the current iteration such that operations of the execution device are controlled according to the iterative action selection policy of the action in the state of the execution device in the current iteration for completing the task.

9. The system of claim 8 , the operations further comprising, in response to determining that the convergence condition is met, controlling the operations of the execution device according to the iterative action selection policy of the action in the state of the execution device in the current iteration for completing the task.

10. The system of claim 8 , wherein computing a regret value of the action in the state of the execution device in the previous iteration based on the iterative action selection policy of the action in the state of the execution device in the previous iteration comprises computing the regret value of the action in the state of the execution device in the previous iteration based on the iterative action selection policy of the action in the state of the execution device in the previous iteration but not any regret value of the action in the state of the execution device in any iteration prior to the previous iteration.

11. The system of claim 8 , wherein computing a regret value of the action in the state of the execution device in the previous iteration based on the iterative action selection policy of the action in the state of the execution device in the previous iteration comprises computing the regret value of the action in the state of the execution device in the previous iteration based on a difference between a counterfactual value of the action in the state of the execution device and a counterfactual value of the state of the execution device, wherein the counterfactual value of the action in the state of the execution device and the counterfactual value of the state of the execution device are computed by recursively traversing a game tree that represents the environment that includes the execution device and the one or more other devices in the previous iteration.

12. The system of claim 8 , wherein the weighted sum of the iterative action selection policy of the action in the state of the execution device in the previous iteration and the incremental action selection policy of the action in the state of the execution device in the current iteration comprises a sum of:

the iterative action selection policy of the action in the state of the execution device in the previous iteration scaled by a first learning rate in the current iteration; and

the incremental action selection policy of the action in the state of the execution device in the current iteration scaled by a second learning rate in the current iteration, and

wherein the first learning rate approaches 1 as the number of iterations, t, approaches infinity, and the second learning rate approaches 0 as t approaches infinity.

13. The system of claim 12 , wherein the first learning rate is (t−1)/t, and the second learning rate is 1/t.

14. The system of claim 8 , the operations further comprising initializing the iterative action selection policy of the action in the state of the execution device in the first iteration based on a uniform random action selection policy.

15. A non-transitory, computer-readable storage medium storing one or more instructions executable by a computer system to perform operations for generating an action selection policy for completing a task in an environment that includes an execution device and one or more other devices, the operations comprising:

at each of a plurality of iterations and for each action among a plurality of possible actions in a state of the execution device in a current iteration, wherein the state of the execution device results from a history of actions taken by the execution device,

identifying an iterative action selection policy of the action in the state of the execution device in a previous iteration, wherein the iterative action selection policy of the action in the state of the execution device in the previous iteration represents a probability of the action taken by the execution device in the state in the previous iteration;

computing a regret value of the action in the state of the execution device in the previous iteration based on the iterative action selection policy of the action in the state of the execution device in the previous iteration, wherein the regret value of the action in the state of the execution device represents a difference between a gain of the execution device attributed to the action towards completing the task after taking the action in the state and a gain of the execution device in the state attributed to the state towards completing the task;

computing an incremental action selection policy of the action in the state of the execution device in the current iteration based on the regret value of the action in the state of the execution device in the previous iteration but not any regret value of the action in the state of the execution device in any iteration prior to the previous iteration;

computing an iterative action selection policy of the action in the state of the execution device in the current iteration based on a weighted sum of the iterative action selection policy of the action in the state of the execution device in the previous iteration and the incremental action selection policy of the action in the state of the execution device in the current iteration;

storing the iterative action selection policy of the action in the state of the execution device in the current iteration; and

in response to determining that a convergence condition is met, outputting the iterative action selection policy of the action in the state of the execution device in the current iteration such that operations of the execution device are controlled according to the iterative action selection policy of the action in the state of the execution device in the current iteration for completing the task.

16. The non-transitory, computer-readable storage medium of claim 15 , the operations further comprising, in response to determining that the convergence condition is met, controlling the operations of the execution device according to the iterative action selection policy of the action in the state of the execution device in the current iteration for completing the task.

17. The non-transitory, computer-readable storage medium of claim 15 , wherein computing a regret value of the action in the state of the execution device in the previous iteration based on the iterative action selection policy of the action in the state of the execution device in the previous iteration comprises computing the regret value of the action in the state of the execution device in the previous iteration based on the iterative action selection policy of the action in the state of the execution device in the previous iteration but not any regret value of the action in the state of the execution device in any iteration prior to the previous iteration.

18. The non-transitory, computer-readable storage medium of claim 15 , wherein computing a regret value of the action in the state of the execution device in the previous iteration based on the iterative action selection policy of the action in the state of the execution device in the previous iteration comprises computing the regret value of the action in the state of the execution device in the previous iteration based on a difference between a counterfactual value of the action in the state of the execution device and a counterfactual value of the state of the execution device, wherein the counterfactual value of the action in the state of the execution device and the counterfactual value of the state of the execution device are computed by recursively traversing a game tree that represents the environment that includes the execution device and the one or more other devices in the previous iteration.

19. The non-transitory, computer-readable storage medium of claim 15 , wherein the weighted sum of the iterative action selection policy of the action in the state of the execution device in the previous iteration and the incremental action selection policy of the action in the state of the execution device in the current iteration comprises a sum of:

the iterative action selection policy of the action in the state of the execution device in the previous iteration scaled by a first learning rate in the current iteration; and

the incremental action selection policy of the action in the state of the execution device in the current iteration scaled by a second learning rate in the current iteration, and

wherein the first learning rate approaches 1 as the number of iterations, t, approaches infinity, and the second learning rate approaches 0 as t approaches infinity.

20. The non-transitory, computer-readable storage medium of claim 19 , wherein the first learning rate is (t−1)/t, and the second learning rate is 1/t.

21. The non-transitory, computer-readable storage medium of claim 15 , the operations further comprising initializing the iterative action selection policy of the action in the state of the execution device in the first iteration based on a uniform random action selection policy.