Solving sparse reward tasks using self-balancing shaped rewards

1. A method for training a goal-oriented task module implemented on one or more hardware processors for performing a reward learning task, comprising:

randomly selecting, by the one or more hardware processors, a start state and a goal state corresponding to a training task for the goal-oriented task module comprising one or more neural networks;

traversing, by the one or more hardware processors, a first trajectory for moving from the start state toward the goal state, the first trajectory ending at a first terminal state;

traversing, by the one or more hardware processors, a second trajectory for moving from the start state toward the goal state, the second trajectory ending at a second terminal state, the first terminal state being closer to the goal state than the second terminal state is to the goal state;

updating a first reward for the first trajectory using a self-balancing reward function based on the second terminal state;

updating a second reward for the second trajectory using the self-balancing reward function based on the first terminal state;

determining, by the one or more hardware processors, a gradient for the goal-oriented task module, wherein the second trajectory contributes to the determination of the gradient and wherein the first trajectory contributes to the determination of the gradient when at least one of the following is true: (i) the first terminal state and the second terminal state are within a first threshold distance of each other, and (ii) the first terminal state is within a second threshold distance from the goal state;

training the goal-oriented task module by updating and storing one or more parameters of the goal-oriented task module at a memory based on the gradient;

generating, by the trained goal-oriented task module, in response to an input indicating an initialized start state and a desired goal state of a goal-oriented task, a series of operations; and

outputting, via a communication interface, the series of operations to an execution system that executes the series of operations to move a system state from the initialized start state to the desired goal state.

2. The method of claim 1 , wherein the first terminal state is an anti-goal for the second trajectory.

3. The method of claim 1 , wherein the first terminal state provides an estimate of a local optimum.

4. The method of claim 1 , wherein updating the first reward for the first trajectory using the self-balancing reward function comprises rewarding the first trajectory when the first terminal state is within the second threshold distance of the goal state.

5. The method of claim 1 , wherein updating the first reward for the first trajectory using the self-balancing reward function comprises:

penalizing the first trajectory based on a distance between the first terminal state and the goal state; and

rewarding the first trajectory based on how close the first terminal state is to the second terminal state.

6. The method of claim 1 , further comprising iteratively determining the first threshold distance.

7. The method of claim 1 , wherein traversing the first trajectory comprises iterating along the first trajectory for a configurable number of iterations.

8. The method of claim 1 , wherein traversing the first trajectory comprises iterating along the first trajectory until the first trajectory converges at the first terminal state.

9. The method of claim 1 , further comprising selecting multiple start and goal state pairs and evaluating multiple first and second trajectory pairs before updating the one or more parameters of the goal-oriented task module.

10. A non-transitory machine-readable medium comprising executable code for training a goal-oriented task module for performing a reward learning task, which when executed by one or more processors associated with a computing device are adapted to cause the one or more processors to perform a method for training a goal-oriented task module, comprising:

randomly selecting, by the one or more hardware processors, a start state and a goal state corresponding to a training task for the goal-oriented task module comprising one or more neural networks;

traversing, by the one or more hardware processors, a first trajectory for moving from the start state toward the goal state, the first trajectory ending at a first terminal state;

traversing, by the one or more hardware processors, a second trajectory for moving from the start state toward the goal state, the second trajectory ending at a second terminal state, the first terminal state being closer to the goal state than the second terminal state is to the goal state;

updating a first reward for the first trajectory using a self-balancing reward function based on the second terminal state;

updating a second reward for the second trajectory using the self-balancing reward function based on the first terminal state;

determining, by the one or more hardware processors, a gradient for the goal-oriented task module, wherein the second trajectory contributes to the determination of the gradient and wherein the first trajectory contributes to the determination of the gradient when at least one of the following is true: (i) the first terminal state and the second terminal state are within a first threshold distance of each other, and (ii) the first terminal state is within a second threshold distance from the goal state;

training the goal-oriented task module by updating and storing one or more parameters of the goal-oriented task module at a memory based on the gradient;

generating, by the trained goal-oriented task module, in response to an input indicating an initialized start state and a desired goal state of a goal-oriented task, a series of operations; and

outputting, via a communication interface, the series of operations to an execution system that executes the series of operations to move a system state from the initialized start state to the desired goal state.

11. The non-transitory machine-readable medium of claim 10 , wherein the first terminal state is an anti-goal for the second rollout.

12. The non-transitory machine-readable medium of claim 10 , wherein the first terminal state provides an estimate of a local optimum.

13. The non-transitory machine-readable medium of claim 10 , wherein updating the first reward for the first rollout using the self-balancing reward function comprises rewarding the first rollout when the first terminal state is within the second threshold distance of the goal state.

14. The non-transitory machine-readable medium of claim 10 , wherein updating the first reward for the first rollout using the self-balancing reward function comprises:

penalizing the first rollout based on a distance between the first terminal state and the goal state; and

rewarding the first rollout based on how close the first terminal state is to the second terminal state.

15. The non-transitory machine-readable medium of claim 10 , further comprising iteratively determining the first threshold distance.

16. The non-transitory machine-readable medium of claim 10 , wherein traversing the first rollout comprises:

iterating along the first rollout for a configurable number of iterations; or

iterating along the first rollout until the first rollout converges at the first terminal state.

17. A device comprising:

a memory storing parameters of a goal-oriented task module and executable code for training the goal-oriented task module; and

one or more processors configured to execute the executable code to perform a method of training a goal-oriented task module for performing a reward learning task, the method comprising:

randomly selecting a start state and a goal state corresponding to a training task for the goal-oriented task module comprising one or more neural networks;

traversing a first trajectory for moving from the start state toward the goal state, the first trajectory ending at a first terminal state;

traversing a second trajectory for moving from the start state toward the goal state, the second trajectory ending at a second terminal state, the first terminal state being closer to the goal state than the second terminal state is to the goal state;

updating a first reward for the first trajectory using a self-balancing reward function based on the second terminal state;

updating a second reward for the second trajectory using the self-balancing reward function based on the first terminal state;

determining a gradient for the goal-oriented task module, wherein the second trajectory contributes to the determination of the gradient and wherein the first trajectory contributes to the determination of the gradient when at least one of the following is true: (i) the first terminal state and the second terminal state are within a first threshold distance of each other, and (ii) the first terminal state is within a second threshold distance from the goal state;

training the goal-oriented task module by updating and storing one or more parameters of the goal-oriented task module at the memory based on the gradient; and

generating, by the trained goal-oriented task module, in response to an input indicating an initialized start state and a desired goal state of a goal-oriented task, a series of operations; and

a communication interface that outputs the series of operations to an execution system that executes the series of operations to move a system state from the initialized start state to the desired goal state.

18. The device of claim 17 , wherein to update the first reward for the first trajectory using the self-balancing reward function, the one or more processors are configured to reward the first trajectory when the first terminal state is within the second threshold distance of the goal state.

19. The device of claim 17 , wherein to update the first reward for the first trajectory using the self-balancing reward function, the one or more processors are configured to:

penalizing the first trajectory based on a distance between the first terminal state and the goal state; and

rewarding the first trajectory based on how close the first terminal state is to the second terminal state.

20. The device of claim 17 , wherein to traverse the first trajectory, the goal-oriented task module is configured to:

iterate along the first trajectory for a configurable number of iterations; or

iterate along the first trajectory until the first trajectory converges at the first terminal state.