Training neural networks using a prioritized experience memory

1. A method for controlling an agent in an environment to perform a task, the method comprising:

receiving a current observation characterizing a current state of the environment;

processing the current observation using a neural network to generate an output that specifies an action to be performed by the agent in response to the current observation, wherein the neural network has been trained through reinforcement learning to determine trained values of parameters of the neural network using a plurality of pieces of selected experience data selected from a prioritized experience memory that stored, during the training of the neural network through reinforcement learning, a plurality of pieces of experience data in association with expected learning progress measures,

wherein each piece of experience data is a training tuple that comprises a training current observation characterizing a training current state of the environment, and a training current action performed by the agent in response to the training current observation, and wherein, for each piece of experience data, a respective value of an expected learning progress measure that is stored in association with the piece of experience data in the prioritized experience memory is derived from a result of a preceding time that values of the parameters of the neural network were updated using the piece of experience data during the training, and

wherein, during the training, the plurality of pieces of selected experience data were selected from the prioritized experience memory based on the respective values of the expected learning progress measures that are stored in association with the plurality of pieces of experience data in the prioritized experience memory; and

causing the agent to perform the action specified by the output in response to the current observation.

2. The method of claim 1 , wherein the agent is a mechanical agent, and the environment is a real-world environment.

3. The method of claim 2 , wherein the mechanical agent comprises a robot, an autonomous, or semi-autonomous vehicle, and wherein causing the agent to perform the action comprises generating control inputs to control the mechanical agent in the real-world environment.

4. The method of claim 1 , wherein the agent is an electronic agent, and the environment is a simulated environment.

5. The method of claim 4 , wherein causing the agent to perform the action comprises generating control inputs to control the electronic agent in the simulated environment.

6. The method of claim 1 , wherein the agent performing the action causes the environment to transition into a new state.

7. The method of claim 1 , wherein causing the agent to perform the action comprises:

receiving a numeric value defining a reward as a consequence of the agent performing the action.

8. The method of claim 1 , wherein causing the agent to perform the action comprises:

selecting, in accordance with the output of the neural network, the action from a predetermined set of actions.

9. The method of claim 8 , wherein the output of the neural network comprises a respective Q value for each action from the predetermined set of actions, and wherein selecting the action from the predetermined set of actions comprises selecting the action that has a highest Q value among the predetermined set of actions.

10. The method of claim 1 , wherein for each piece of experience data, the expected amount of progress is dependent on a temporal difference learning error determined in the training of the neural network through reinforcement learning using the piece of experience data.

11. The method of claim 1 , further comprising, during the training of the neural network through reinforcement learning:

deleting one or more pieces of experience data from the replay memory.

12. A system comprising one or more computers and one or more storage devices storing instructions that are operable, when executed by the one or more computers, to cause the one or more computers to perform operations for controlling an agent in an environment to perform a task, the operations comprising:

receiving a current observation characterizing a current state of the environment;

processing the current observation using a neural network to generate an output that specifies an action to be performed by the agent in response to the current observation, wherein the neural network has been trained through reinforcement learning to determine trained values of parameters of the neural network using a plurality of pieces of selected experience data selected from a prioritized experience memory that stored, during the training of the neural network through reinforcement learning, a plurality of pieces of experience data in association with expected learning progress measures,

wherein each piece of experience data is a training tuple that comprises a training current observation characterizing a training current state of the environment, and a training current action performed by the agent in response to the training current observation, and wherein, for each piece of experience data, a respective value of an expected learning progress measure that is stored in association with the piece of experience data in the prioritized experience memory is derived from a result of a preceding time that values of the parameters of the neural network were updated using the piece of experience data during the training, and

wherein, during the training, the plurality of pieces of selected experience data were selected from the prioritized experience memory based on the respective values of the expected learning progress measures that are stored in association with the plurality of pieces of experience data in the prioritized experience memory; and

causing the agent to perform the action specified by the output in response to the current observation.

13. The system of claim 12 , wherein the agent is a mechanical agent, and the environment is a real-world environment.

14. The system of claim 13 , wherein the mechanical agent comprises a robot, an autonomous, or semi-autonomous vehicle, and wherein causing the agent to perform the action comprises generating control inputs to control the mechanical agent in the real-world environment.

15. The system of claim 12 , wherein the agent is an electronic agent, and the environment is a simulated environment.

16. The system of claim 15 , wherein causing the agent to perform the action comprises generating control inputs to control the electronic agent in the simulated environment.

17. The system of claim 12 , wherein the agent performing the action causes the environment to transition into a new state.

18. The system of claim 12 , wherein causing the agent to perform the action comprises:

receiving a numeric value defining a reward as a consequence of the agent performing the action.

19. The system of claim 12 , wherein causing the agent to perform the action comprises:

selecting, in accordance with the output of the neural network, the action from a predetermined set of actions.

20. A non-transitory computer storage medium encoded with instructions that, when executed by one or more computers, cause the one or more computers to perform operations for controlling an agent in an environment to perform a task, the operations comprising:

receiving a current observation characterizing a current state of the environment;

processing the current observation using a neural network to generate an output that specifies an action to be performed by the agent in response to the current observation, wherein the neural network has been trained through reinforcement learning to determine trained values of parameters of the neural network using a plurality of pieces of selected experience data selected from a prioritized experience memory that stored, during the training of the neural network through reinforcement learning, a plurality of pieces of experience data in association with expected learning progress measures,

wherein each piece of experience data is a training tuple that comprises a training current observation characterizing a training current state of the environment, and a training current action performed by the agent in response to the training current observation, and wherein, for each piece of experience data, a respective value of an expected learning progress measure that is stored in association with the piece of experience data in the prioritized experience memory is derived from a result of a preceding time that values of the parameters of the neural network were updated using the piece of experience data during the training, and

wherein, during the training, the plurality of pieces of selected experience data were selected from the prioritized experience memory based on the respective values of the expected learning progress measures that are stored in association with the plurality of pieces of experience data in the prioritized experience memory; and

causing the agent to perform the action specified by the output in response to the current observation.