Unsupervised control using learned rewards

1. A method performed by one or more computers, the method comprising:

selecting actions to be performed by an agent that interacts with an environment by performing actions from a predetermined set of actions, comprising, at each of a plurality of time steps:

receiving an observation characterizing a current state of the environment at the time step and an observation characterizing a goal state of the environment at the time step;

processing the observation characterizing the current state of the environment at the time step and the observation characterizing the goal state of the environment at the time step using an action selection neural network to generate an output comprising an action selection policy; and

selecting an action from the predetermined set of actions to be performed by the agent at the time step using the action selection policy;

wherein the action selection neural network has been trained by operations comprising, for each of a plurality of training time steps:

receiving an observation characterizing a current state of the environment at the training time step and an observation characterizing a goal state of the environment at the training time step;

processing the observation characterizing the current state of the environment at the training time step and the observation characterizing the goal state of the environment at the training time step using an embedding neural network, wherein the embedding neural network is configured to process an observation to generate an embedded representation of the observation; and

generating a reward for the training time step using the embedded representation of the observation characterizing the current state of the environment at the training time step and the embedded representation of the observation characterizing the goal state of the environment at the training time step; and

training the action selection neural network based on the reward for the training time step using a reinforcement learning technique.

2. The method of claim 1 , wherein generating the reward for the training time step comprises:

determining a similarity between the observation characterizing the current state of the environment and the observation characterizing the goal state of the environment using: (i) the embedded representation of the observation characterizing the current state of the environment, and (ii) the embedded representation of the observation characterizing the goal state of the environment; and

determining the reward for the training time step based on the similarity between the observation characterizing the current state of the environment and the observation characterizing the goal state of the environment.

3. The method of claim 2 , wherein determining the similarity between the observation characterizing the current state of the environment and the observation characterizing the goal state of the environment comprises:

obtaining one or more decoy observations, wherein each decoy observation characterizes a past state of the environment;

for each decoy observation, generate an embedded representation of the decoy observation by processing the decoy observation using the embedding neural network;

determine the similarity between the observation characterizing the current state of the environment and the observation characterizing the goal state of the environment using: (i) the embedded representation of the observation characterizing the current state of the environment, (ii) the embedded representation of the observation characterizing the goal state of the environment, and (iii) the embedded representations of the decoy observations.

4. The method of claim 3 , wherein determining the similarity between the observation characterizing the current state of the environment and the observation characterizing the goal state of the environment comprises:

for each decoy observation, determining a similarity metric between the embedded representation of the observation characterizing the current state of the environment and the embedded representation of the decoy observation;

determining a similarity metric between the embedded representation of the observation characterizing the current state of the environment and the embedded representation of the observation characterizing the goal state of the environment; and

determining the similarity between the observation characterizing the current state of the environment and the observation characterizing the goal state of the environment using the determined similarity metrics.

5. The method of claim 4 , wherein the similarity metric is an inner product.

6. The method of claim 2 , wherein determining the similarity between the observation characterizing the current state of the environment and the observation characterizing the goal state of the environment comprises:

evaluating a similarity metric between the embedded representation of the observation characterizing the current state of the environment and the embedded representation of the observation characterizing the goal state of the environment.

7. The method of claim 2 , wherein the training further comprises:

determine a gradient of a loss function that depends on the determined similarity between the observation characterizing the current state of the environment and the observation characterizing the goal state of the environment; and

adjust current values of embedding neural network parameters using the gradient.

8. The method of claim 1 , wherein the training further comprises, at one or more of the plurality of training time steps:

determining the observation characterizing the goal state of the environment by sampling from a goal buffer, wherein the goal buffer comprises a plurality of past observations.

9. The method of claim 8 , wherein the training further comprises, at one or more of the plurality of training time steps:

updating the goal buffer by replacing a randomly selected observation from the goal buffer by a current observation with a predetermined probability.

10. The method of claim 8 , wherein the training further comprises, at one or more of the plurality of training time steps:

randomly selecting a candidate observation from the goal buffer;

determining whether to replace the candidate observation by a current observation based on: (i) an aggregate similarity measure between the current observation and the observations included in the goal buffer, and (ii) an aggregate similarity measure between the candidate observation and the observations included in the goal buffer.

11. The method of claim 1 , wherein the action selection neural network is a Q neural network and the action selection policy comprises a respective Q value for each action in the predetermined set of actions.

12. The method of claim 11 , wherein at each of the plurality of time steps, selecting an action comprises selecting an action with a highest Q value.

13. The method of claim 1 , wherein at each of the plurality of time steps, selecting an action comprises selecting the action in accordance with an exploration policy.

14. The method of claim 1 , wherein the action selection neural network and the embedding neural network are jointly trained.

15. The method of claim 1 , wherein at each of the plurality of time steps, the observation characterizing the current state of the environment at the time step comprises one or more of: image data, video data, point cloud data, temperature data, motor current data, voltage data, or power data.

16. The method of claim 1 , wherein the agent selects actions to be performed in a real-world environment, wherein the real-world environment comprises a data center environment, a power distribution environment, a water distribution environment, or a manufacturing plant environment.

17. The method of claim 1 , further comprising, for each of one or more of the plurality of training time steps:

receiving an external reward for the training time step;

wherein the reward for the training time step is generated based at least in part on the external reward for the training time step.

18. The method of claim 1 , wherein the embedded representation of the observation characterizing the goal state of the environment is generated once at a first training time step and reused at each subsequent training time step.

19. A system comprising:

one or more computers; and

one or more storage devices communicatively coupled to the one or more computers, wherein the one or more storage devices store instructions that, when executed by the one or more computers, cause the one or more computers to perform operations comprising:

selecting actions to be performed by an agent that interacts with an environment by performing actions from a predetermined set of actions, comprising, at each of a plurality of time steps:

receiving an observation characterizing a current state of the environment at the time step and an observation characterizing a goal state of the environment at the time step;

processing the observation characterizing the current state of the environment at the time step and the observation characterizing the goal state of the environment at the time step using an action selection neural network to generate an output comprising an action selection policy; and

selecting an action from the predetermined set of actions to be performed by the agent at the time step using the action selection policy;

wherein the action selection neural network has been trained by operations comprising, for each of a plurality of training time steps:

receiving an observation characterizing a current state of the environment at the training time step and an observation characterizing a goal state of the environment at the training time step;

processing the observation characterizing the current state of the environment at the training time step and the observation characterizing the goal state of the environment at the training time step using an embedding neural network, wherein the embedding neural network is configured to process an observation to generate an embedded representation of the observation; and

generating a reward for the training time step using the embedded representation of the observation characterizing the current state of the environment at the training time step and the embedded representation of the observation characterizing the goal state of the environment at the training time step; and

training the action selection neural network based on the reward for the training time step using a reinforcement learning technique.

20. One or more non-transitory computer storage media storing instructions that when executed by one or more computers cause the one or more computers to perform operations comprising:

selecting actions to be performed by an agent that interacts with an environment by performing actions from a predetermined set of actions, comprising, at each of a plurality of time steps:

receiving an observation characterizing a current state of the environment at the time step and an observation characterizing a goal state of the environment at the time step;

processing the observation characterizing the current state of the environment at the time step and the observation characterizing the goal state of the environment at the time step using an action selection neural network to generate an output comprising an action selection policy; and

selecting an action from the predetermined set of actions to be performed by the agent at the time step using the action selection policy;

wherein the action selection neural network has been trained by operations comprising, for each of a plurality of training time steps:

receiving an observation characterizing a current state of the environment at the training time step and an observation characterizing a goal state of the environment at the training time step;

processing the observation characterizing the current state of the environment at the training time step and the observation characterizing the goal state of the environment at the training time step using an embedding neural network, wherein the embedding neural network is configured to process an observation to generate an embedded representation of the observation; and

generating a reward for the training time step using the embedded representation of the observation characterizing the current state of the environment at the training time step and the embedded representation of the observation characterizing the goal state of the environment at the training time step; and

training the action selection neural network based on the reward for the training time step using a reinforcement learning technique.