Training action selection neural networks using apprenticeship
An off-policy reinforcement learning actor-critic neural network system configured to select actions from a continuous action space to be performed by an agent interacting with an environment to perform a task. An observation defines environment state data and reward data. The system has an actor neural network which learns a policy function mapping the state data to action data. A critic neural network learns an action-value (Q) function. A replay buffer stores tuples of the state data, the action data, the reward data and new state data. The replay buffer also includes demonstration transition data comprising a set of the tuples from a demonstration of the task within the environment. The neural network system is configured to train the actor neural network and the critic neural network off-policy using stored tuples from the replay buffer comprising tuples both from operation of the system and from the demonstration transition data.
1. A computer-implemented method for controlling an agent interacting with an environment to perform a task, comprising:
receiving an observation comprising input state data characterizing a state of the environment;
processing the input state data using an actor neural network to generate an output including output action data defining an output action; and
controlling the agent to perform the output action;
wherein the actor neural network has been trained jointly with a critic neural network that is configured to define a value function that generates an error signal based on an input comprising action data defining an action, state data characterizing the state of the environment, and return data derived from reward data representing a reward from the action performed in a training process; wherein the training process has been performed based on training data including data from a demonstration of the task being performed by an expert agent within the environment, wherein the data from the demonstration includes demonstration transition data for a series of demonstration transitions including demonstration examples of the state data, the action data, the reward data, and new state data representing a new state that have been generated as a result of the expert agent interacting with the environment; wherein during the training process, the actor neural network has been used to operate on the environment to generate operation transition data comprising operational examples of the state data, the action data, the reward data and the new state data, and the actor neural network and the critic neural network have been trained off-policy using the error signal and using stored tuples sampled from a replay buffer comprising tuples from both the operation transition data and the demonstration transition data.
2. The method of claim 1 , wherein the reward comprises a sparse reward that has a plurality of discrete values dependent upon the state of the environment.
3. The method of claim 1 , wherein only a minority subset of states of the environment provide the reward.
4. The method of claim 1 , wherein during the training process, the stored tuples have been sampled from the replay buffer for training the actor neural network and the critic neural network according to a sampling probability that prioritizes sampling of tuples of the demonstration examples.
5. The method of claim 1 , wherein the return data comprises a combination of the reward data and values from the critic neural network obtained from an (n-1)-step forward rollout of actions selected using the actor neural network; and wherein the training process employs at least two different values of n to train the network.
6. The method of claim 1 , wherein during the training process, the critic neural network has been trained using return data which comprises a mix of 1-step and n-step returns.
7. The method of claim 1 , wherein during the training process:
weights of a learning critic neural network have been updated off-policy using the error signal determined from a target actor neural network and a target critic neural network, and updating weights of a learning actor neural network using a deterministic policy gradient comprising a product of a gradient of the output of the learning critic neural network and a gradient of the output of the learning actor neural network evaluated using the stored tuples of both the operation transition data and the demonstration transition data; wherein:
weights of the target actor neural network have been updated at intervals using the learning actor neural network; and weights of the target critic neural network have been updated at intervals using the learning critic neural network.
8. The method of claim 1 , wherein the training data comprises kinesthetic teaching data from manipulation of a mechanical system.
9. The method of claim 1 , further comprising a safety controller to impose safety or other constraints on the action data.
10. A computer-implemented system for controlling an agent interacting with an environment to perform a task, comprising:
one or more processors; and
one or more storage devices storing instructions that when executed by the one or more computers, cause the one or more processors to perform operations comprising:
receiving an observation comprising input state data characterizing a state of the environment;
processing the input state data using an actor neural network to generate an output including output action data defining an output action; and
controlling the agent to perform the output action;
wherein the actor neural network has been trained jointly with a critic neural network that is configured to define a value function that generates an error signal based on an input comprising action data defining an action, state data characterizing the state of the environment, and return data derived from reward data representing a reward from the action performed in a training process; wherein the training process has been performed based on training data including data from a demonstration of the task being performed by an expert agent within the environment, wherein the data from the demonstration includes demonstration transition data for a series of demonstration transitions including demonstration examples of the state data, the action data, the reward data, and new state data representing a new state that have been generated as a result of the expert agent interacting with the environment; wherein during the training process, the actor neural network has been used to operate on the environment to generate operation transition data comprising operational examples of the state data, the action data, the reward data and the new state data, and the actor neural network and the critic neural network have been trained off-policy using the error signal and using stored tuples sampled from a replay buffer comprising tuples from both the operation transition data and the demonstration transition data.
11. The computer-implemented system of claim 10 , wherein the reward comprises a sparse reward that has a plurality of discrete values dependent upon the state of the environment.
12. The computer-implemented system of claim 10 , wherein only a minority subset of states of the environment provide the reward.
13. The computer-implemented system of claim 10 , wherein during the training process,
the stored tuples have been sampled from the replay buffer for training the actor neural network and the critic neural network according to a sampling probability that prioritizes sampling of tuples of the demonstration examples.
14. The computer-implemented system of claim 10 , wherein the return data comprises a combination of the reward data and values from the critic neural network obtained from an (n-1)-step forward rollout of actions selected using the actor neural network; and wherein the training process employs at least two different values of n to train the network.
15. The computer-implemented system of claim 10 , wherein during the training process, the critic neural network has been trained using return data which comprises a mix of 1-step and n-step returns.
16. One or more non-transitory computer-readable storage media storing instructions 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, the operations comprising:
receiving an observation comprising input state data characterizing a state of the environment;
processing the input state data using an actor neural network to generate an output including output action data defining an output action; and
controlling the agent to perform the output action;
wherein the actor neural network has been trained jointly with a critic neural network that is configured to define a value function that generates an error signal based on an input comprising action data defining an action, state data characterizing the state of the environment, and return data derived from reward data representing a reward from the action performed in a training process; wherein the training process has been performed based on training data including data from a demonstration of the task being performed by an expert agent within the environment, wherein the data from the demonstration includes demonstration transition data for a series of demonstration transitions including demonstration examples of the state data, the action data, the reward data, and new state data representing a new state that have been generated as a result of the expert agent interacting with the environment; wherein during the training process, the actor neural network has been used to operate on the environment to generate operation transition data comprising operational examples of the state data, the action data, the reward data and the new state data, and the actor neural network and the critic neural network have been trained off-policy using the error signal and using stored tuples sampled from a replay buffer comprising tuples from both the operation transition data and the demonstration transition data.
17. The one or more non-transitory computer-readable storage media of claim 16 , wherein the reward comprises a sparse reward that has a plurality of discrete values dependent upon the state of the environment.
18. The one or more non-transitory computer-readable storage media of claim 16 , wherein only a minority subset of states of the environment provide the reward.
19. The one or more non-transitory computer-readable storage media of claim 16 , wherein during the training process, the stored tuples have been sampled from the replay buffer for training the actor neural network and the critic neural network according to a sampling probability that prioritizes sampling of tuples of the demonstration transition examples.
20. The one or more non-transitory computer-readable storage media of claim 16 , wherein the return data comprises a combination of the reward data and values from the critic neural network obtained from an (n-1)-step forward rollout of actions selected using the actor neural network; and wherein the training process employs at least two different values of n to train the network.