CONTROLLING AGENTS BY TRANSFERRING SUCCESSOR FEATURES TO NEW TASKS

1 . A method performed by one or more computers and for controlling an agent interacting with an environment to perform a task episode of a new task, the method comprising, at each of a plurality of current time steps in the task episode of the new task:

obtaining an observation characterizing a current state of the environment at the time step;

obtaining a learned task encoding of the new task;

generating, from the current observation and the learned task encoding for the new task, a dynamic transfer query vector for the current time step;

for each of a plurality of training tasks and for each action in a candidate set of actions:

determining, from the current observation, an action-specific policy vector for the candidate action for the training task given the current state of the environment at the time step; and

 determining, based on a similarity between the dynamic transfer query vector for the current time step and the action-specific policy vector, a quality score for the candidate action and for the training task; and

selecting an action from the candidate set using the respective quality scores for the candidate actions for the training tasks.

2 . The method of claim 1 , further comprising:

generating the learned task encoding of the new task for the current time step using a task encoding neural network.

3 . The method of claim 2 , wherein generating the learned task encoding of the new task comprises processing data characterizing the new task using the task encoding neural network to generate a vector describing features associated with the new task.

4 . The method of claim 3 , wherein the data characterizing the new task comprises at last one of text describing the new task or an observation characterizing a goal state for the new task.

5 . The method of claim 1 , wherein selecting an action to be performed by the agent based on the respective quality scores for the candidate actions for the plurality of training tasks comprises:

identifying a candidate action that has a highest quality score; and

selecting the identified action.

6 . The method of claim 1 , wherein determining, based on a similarity between the dynamic transfer query vector for the current time step and the action-specific policy vector, a quality score for the candidate action and for the training task comprises:

computing an inner product between (i) the action-specific policy vector for the candidate action for the training task and (ii) the dynamic transfer query vector.

7 . The method of claim 1 , wherein generating, from the current observation and the learned task encoding for the new task, a dynamic transfer query vector for the current time step comprises:

processing the encoded representation of the observation and the learned task encoding of the new task using a new task state representation neural network to generate a new state vector that represents a state of the new task as of the current time step; and

generating, from the new state vector, the dynamic transfer query vector for the current time step.

8 . The method of claim 1 , wherein the dynamic transfer query vector is a linear combination of respective training task encodings of each of the plurality of training tasks.

9 . The method of claim 1 , wherein generating, from the new state vector, the dynamic transfer query vector comprises:

processing the new state vector using a multilayer perceptron neural network to generate an output defining a linear combination of respective training task encodings of each of the plurality of training tasks.

10 . The method of claim 9 , wherein generating the dynamic transfer query vector further comprises:

sampling, using the output, coefficients for the respective training task encodings; and

computing a weighted sum of the respective training task encodings in an accordance with the sampled coefficients to generate the dynamic transfer query vector.

11 . The method of claim 1 , wherein generating an action-specific policy vector for the candidate action for the training task comprises:

processing (i) the candidate action and (ii) the encoded representation of the observation using a successor feature neural network to generate successor features for the training task and associated with the candidate action; and

generating an action-specific policy vector comprising the successor features.

12 . The method of claim 11 , wherein the successor feature is modelled as a probability mass function over a discretized range of continuous values.

13 . The method of claim 11 , wherein the successor feature is a value function representation of a discounted sum of future state features associated with the candidate action.

14 . The method of claim 11 , wherein the successor feature neural network is pre-trained on the plurality of training tasks.

15 . The method of claim 14 , the method further comprising jointly training the task encoder neural network and the state feature neural network to optimize an objective function that includes a regression loss term.

16 . The method of claim 14 , the method further comprising:

generating a quality score estimate for the training task using the value function representation; and

training the successor feature neural network to maximize future quality score estimates.

17 . The method of claim 11 , wherein the successor feature neural network is shared across the training tasks and wherein processing (i) the candidate action and (ii) the encoded representation of the observation using a successor feature prediction neural network to generate a successor feature associated with the candidate action comprises:

processing (i) the candidate action (ii) the encoded representation of the observation and (iii) a task encoding of the training task using a successor feature prediction neural network to generate a successor feature associated with the candidate action.

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

19 . The method of claim 18 , wherein the agent is a robot.

20 . The method of claim 1 , wherein the training tasks are learned in a simulated environment and the new task is performed in a real-world environment.

21 . A system comprising one or more computers and one or more storage devices storing instruction that when executed by the one or more computers cause the one or more computers to perform operations for controlling an agent interacting with an environment to perform a task episode of a new task, the method comprising, at each of a plurality of current time steps in the task episode of the new task:

obtaining an observation characterizing a current state of the environment at the time step;

obtaining a learned task encoding of the new task;

generating, from the current observation and the learned task encoding for the new task, a dynamic transfer query vector for the current time step;

for each of a plurality of training tasks and for each action in a candidate set of actions:

determining, from the current observation, an action-specific policy vector for the candidate action for the training task given the current state of the environment at the time step; and

determining, based on a similarity between the dynamic transfer query vector for the current time step and the action-specific policy vector, a quality score for the candidate action and for the training task; and

selecting an action from the candidate set using the respective quality scores for the candidate actions for the training tasks.

22 . One or more non-transitory computer storage media storing instruction that when executed by one or more computers cause the one or more computers to perform operations for controlling an agent interacting with an environment to perform a task episode of a new task, the method comprising, at each of a plurality of current time steps in the task episode of the new task:

obtaining an observation characterizing a current state of the environment at the time step;

obtaining a learned task encoding of the new task;

generating, from the current observation and the learned task encoding for the new task, a dynamic transfer query vector for the current time step;

for each of a plurality of training tasks and for each action in a candidate set of actions:

determining, from the current observation, an action-specific policy vector for the candidate action for the training task given the current state of the environment at the time step; and

determining, based on a similarity between the dynamic transfer query vector for the current time step and the action-specific policy vector, a quality score for the candidate action and for the training task; and

selecting an action from the candidate set using the respective quality scores for the candidate actions for the training tasks.