Distributional reinforcement learning
Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for selecting an action to be performed by a reinforcement learning agent interacting with an environment. A current observation characterizing a current state of the environment is received. For each action in a set of multiple actions that can be performed by the agent to interact with the environment, a probability distribution is determined over possible Q returns for the action-current observation pair. For each action, a measure of central tendency of the possible Q returns with respect to the probability distributions for the action-current observation pair is determined. An action to be performed by the agent in response to the current observation is selected using the measures of central tendency.
1. A method performed by one or more data processing apparatus for training a distributional Q network, the method comprising:
obtaining an experience tuple that includes (i) a current training observation, (ii) a current action performed by an agent in response to the current training observation, (iii) a current reward received in response to the agent performing the current action, and (iv) a next training observation characterizing a state that an environment transitioned into as a result of the agent performing the current action;
processing the current training observation and the current action using the distributional Q network and in accordance with current values of network parameters to generate a current network output comprising a plurality of numerical values that collectively define a current probability distribution over possible Q returns for the current action-current training observation pair;
processing a possible next action and the next training observation using a target distributional Q network and in accordance with current values of target network parameters of the distributional Q network to generate a next network output for the action-next training observation pair comprising a plurality of numerical values that collectively define a next probability distribution over possible Q returns for the possible next action-next training observation pair, wherein the target distributional Q network has the same neural network architecture as the distributional Q network but the current values of the target network parameters are different from the current values of the network parameters;
determining a gradient with respect to the network parameters of a loss function that depends on: (i) the current probability distribution over possible Q returns for the current action-current training observation pair, (ii) the next probability distribution over possible Q returns for the possible next action-next training observation pair, and (iii) the current reward, comprising:
determining a respective projected sample update for each of the possible Q returns using the current reward and the possible next action; and
wherein the loss function depends on the projected sample updates for the possible Q returns; and
updating the current values of the network parameters using the gradient.
2. The method of claim 1 , wherein determining a respective projected sample update for each of the possible Q returns using the current reward and the possible next action comprises:
determining a respective sample update for each of the possible Q returns from the current reward; and
determining the respective projected sample update for each of the possible Q returns from the respective sample updates and the next probability distribution for the possible next action-next training observation pair.
3. The method of claim 2 , wherein the respective sample update for each of the possible Q returns is equal to the current reward plus a product of a discount factor and the possible Q return subject to a constraint that the respective sample update not be less than a smallest possible Q return of the plurality of possible Q returns and not be greater than a largest possible Q return of the plurality of possible Q returns.
4. The method of claim 2 , wherein determining the respective projected sample update for each of the possible Q returns from the respective sample updates and the next probability distribution for the possible next action-next training observation pair comprises, for each possible Q return:
distributing a probability for the possible Q return in the next probability distribution for the possible next action-next training observation pair to at least some of the projected sample updates with a strength that is based on, for each projected sample update, a distance between the sample update for the possible Q return and the corresponding possible Q return for the projected sample update.
5. The method of claim 2 , wherein the loss function comprises a Kullback-Leibler divergence between (i) the respective projected sample updates and (ii) the current probability distribution.
6. The method of claim 2 , wherein the possible Q returns are indexed from 0 to N−1, and wherein determining the respective projected sample update comprises, for each particular possible Q return of the plurality of possible Q returns:
remapping the sample update for the particular possible Q return to fall in a range of 0 to N−1;
determining a floor and a ceiling of the remapped sample update;
updating the projected sample update for the possible Q return having an index that matches the floor based on a probability for the particular possible Q return in the next probability distribution for the possible next action-next training observation pair and on a distance between the remapped sample update and the ceiling; and
updating the projected sample update for the possible Q return having an index that matches the ceiling based on a probability for the particular possible Q return in the next probability distribution for the possible next action-next training observation pair and on a distance between the remapped sample update and the floor.
7. The method of claim 6 , wherein the loss function comprises a negative of a sum of, for each possible Q return, the projected sample update for the possible Q return and a logarithm of a probability for the possible Q return under the current probability distribution.
8. The method of claim 1 , further comprising:
periodically updating the values of the target network parameters to match the values of the network parameters.
9. The method claim 1 , further comprising:
updating the current values of the target network parameters by interpolating between the current values of the target network parameters and the updated values of the network parameters.
10. A system comprising:
one or more computers comprising one or more processors; 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 for training a distributional Q network, the operations comprising:
obtaining an experience tuple that includes (i) a current training observation, (ii) a current action performed by an agent in response to the current training observation, (iii) a current reward received in response to the agent performing the current action, and (iv) a next training observation characterizing a state that an environment transitioned into as a result of the agent performing the current action;
processing the current training observation and the current action using the distributional Q network and in accordance with current values of network parameters to generate a current network output comprising a plurality of numerical values that collectively define a current probability distribution over possible Q returns for the current action-current training observation pair;
processing a possible next action and the next training observation using a target distributional Q network and in accordance with current values of target network parameters of the distributional Q network to generate a next network output for the action-next training observation pair comprising a plurality of numerical values that collectively define a next probability distribution over possible Q returns for the possible next action-next training observation pair, wherein the target distributional Q network has the same neural network architecture as the distributional Q network but the current values of the target network parameters are different from the current values of the network parameters;
determining a gradient with respect to the network parameters of a loss function that depends on: (i) the current probability distribution over possible Q returns for the current action-current training observation pair, (ii) the next probability distribution over possible Q returns for the possible next action-next training observation pair, and (iii) the current reward, comprising:
determining a respective projected sample update for each of the possible Q returns using the current reward and the possible next action; and
wherein the loss function depends on the projected sample updates for the possible Q returns; and
updating the current values of the network parameters using the gradient.
11. The system of claim 10 , wherein determining a respective projected sample update for each of the possible Q returns using the current reward and the possible next action comprises:
determining a respective sample update for each of the possible Q returns from the current reward; and
determining the respective projected sample update for each of the possible Q returns from the respective sample updates and the next probability distribution for the possible next action-next training observation pair.
12. The system of claim 11 , wherein the respective sample update for each of the possible Q returns is equal to the current reward plus a product of a discount factor and the possible Q return subject to a constraint that the respective sample update not be less than a smallest possible Q return of the plurality of possible Q returns and not be greater than a largest possible Q return of the plurality of possible Q returns.
13. The system of claim 11 , wherein determining the respective projected sample update for each of the possible Q returns from the respective sample updates and the next probability distribution for the possible next action-next training observation pair comprises, for each possible Q return:
distributing a probability for the possible Q return in the next probability distribution for the possible next action-next training observation pair to at least some of the projected sample updates with a strength that is based on, for each projected sample update, a distance between the sample update for the possible Q return and the corresponding possible Q return for the projected sample update.
14. 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 for training a distributional Q network, the operations comprising:
obtaining an experience tuple that includes (i) a current training observation, (ii) a current action performed by an agent in response to the current training observation, (iii) a current reward received in response to the agent performing the current action, and (iv) a next training observation characterizing a state that an environment transitioned into as a result of the agent performing the current action;
processing the current training observation and the current action using the distributional Q network and in accordance with current values of network parameters to generate a current network output comprising a plurality of numerical values that collectively define a current probability distribution over possible Q returns for the current action-current training observation pair;
processing a possible next action and the next training observation using a target distributional Q network and in accordance with current values of target network parameters of the distributional Q network to generate a next network output for the action-next training observation pair comprising a plurality of numerical values that collectively define a next probability distribution over possible Q returns for the possible next action-next training observation pair, wherein the target distributional Q network has the same neural network architecture as the distributional Q network but the current values of the target network parameters are different from the current values of the network parameters;
determining a gradient with respect to the network parameters of a loss function that depends on: (i) the current probability distribution over possible Q returns for the current action-current training observation pair, (ii) the next probability distribution over possible Q returns for the possible next action-next training observation pair, and (iii) the current reward, comprising:
determining a respective projected sample update for each of the possible Q returns using the current reward and the possible next action; and
wherein the loss function depends on the projected sample updates for the possible Q returns; and
updating the current values of the network parameters using the gradient.
15. The non-transitory computer storage media of claim 14 , wherein determining a respective projected sample update for each of the possible Q returns using the current reward and the possible next action comprises:
determining a respective sample update for each of the possible Q returns from the current reward; and
determining the respective projected sample update for each of the possible Q returns from the respective sample updates and the next probability distribution for the possible next action-next training observation pair.
16. The non-transitory computer storage media of claim 15 , wherein the respective sample update for each of the possible Q returns is equal to the current reward plus a product of a discount factor and the possible Q return subject to a constraint that the respective sample update not be less than a smallest possible Q return of the plurality of possible Q returns and not be greater than a largest possible Q return of the plurality of possible Q returns.
17. The non-transitory computer storage media of claim 15 , wherein determining the respective projected sample update for each of the possible Q returns from the respective sample updates and the next probability distribution for the possible next action-next training observation pair comprises, for each possible Q return:
distributing a probability for the possible Q return in the next probability distribution for the possible next action-next training observation pair to at least some of the projected sample updates with a strength that is based on, for each projected sample update, a distance between the sample update for the possible Q return and the corresponding possible Q return for the projected sample update.
18. The non-transitory computer storage media of claim 15 , wherein the loss function comprises a Kullback-Leibler divergence between (i) the respective projected sample updates and (ii) the current probability distribution.
19. The non-transitory computer storage media of claim 15 , wherein the possible Q returns are indexed from 0 to N−1, and wherein determining the respective projected sample update comprises, for each particular possible Q return of the plurality of possible Q returns:
remapping the sample update for the particular possible Q return to fall in a range of 0 to N−1;
determining a floor and a ceiling of the remapped sample update;
updating the projected sample update for the possible Q return having an index that matches the floor based on a probability for the particular possible Q return in the next probability distribution for the possible next action-next training observation pair and on a distance between the remapped sample update and the ceiling; and
updating the projected sample update for the possible Q return having an index that matches the ceiling based on a probability for the particular possible Q return in the next probability distribution for the possible next action-next training observation pair and on a distance between the remapped sample update and the floor.
20. The non-transitory computer storage media of claim 19 , wherein the loss function comprises a negative of a sum of, for each possible Q return, the projected sample update for the possible Q return and a logarithm of a probability for the possible Q return under the current probability distribution.