Generative neural network systems for generating instruction sequences to control an agent performing a task
Methods, systems, and apparatus for providing a sequence of actions to perform a task. In one aspect, a method comprises: using a policy neural network to, at each of a sequence of time steps, select one or more actions to be performed according to an action selection policy learned by the policy neural network; providing the selected one or more actions to a simulator; implementing the selected one or more actions for the time steps using the simulator to generate a simulator output; discriminating between the simulator output and training data using a discriminator neural network to produce a discriminator output; and updating parameters of the policy recurrent neural network using a reinforcement learning procedure according to a reward signal determined from the discriminator output; and updating parameters of the discriminator neural network according to a difference between the simulator output and the training data.
1 . A system comprising one or more computers and one or more storage devices storing instructions that when executed by the one or more computers cause the one or more computers to perform operations for synthesizing programs for generating images, the operations comprising:
using a policy neural network to, at each of a sequence of time steps, select one or more actions for generating an image to be performed, wherein each action is a respective command to a graphics software configured to generate an image;
controlling the graphics software using the respective commands to cause the graphics software to generate a graphics output comprising an image;
processing the graphics output generated by the graphics software using a discriminator neural network to produce a discriminator output that indicates whether the graphics output is from a set of training data comprising images or was generated by the graphics software; and
training the policy neural network using a reinforcement learning procedure according to a reward signal determined from the discriminator output; and
training the discriminator neural network according to the discriminator output.
2 . The system of claim 1 wherein the graphics software is a non-differentiable graphics engine.
3 . The system of claim 1 wherein the policy neural network has an internal state, the operations further comprising:
determining an estimated return from the internal state using a value function neural network, and
wherein training the policy neural network comprises updating parameters of the policy neural network using the estimated return.
4 . The system of claim 1 wherein training the policy neural network comprises updating parameters of the policy neural network using stored trajectories from an experience buffer that stores trajectories from graphics outputs generated by a plurality of actors.
5 . The system of claim 1 , the operations further comprising storing the graphics output in a reply buffer, and wherein the discriminator neural network is configured to sample from the replay buffer to provide the discriminator output.
6 . The system of claim 1 wherein training the discriminator neural network comprises updating parameters of the discriminator neural network so that the discriminator output is dependent upon a first Wasserstein distance between the graphics output and training data.
7 . The system of claim 1 wherein the reward signal further comprises an auxiliary reward dependent upon the selected one or more actions.
8 . The system of claim 1 wherein the policy neural network has an input to receive a target vector defining a target for the graphics output, and wherein the one or more actions are selected conditionally upon the target vector.
9 . The system of claim 8 , the operations further comprising providing the target vector as the training data for the discriminator neural network.
10 . The system of claim 1 wherein the graphics software comprises a programmable signal generator, and wherein the selected one or more actions for the sequence of time steps comprise programming actions to control the programmable signal generator to provide a graphics output comprising a signal generated according to the programming actions.
11 . A method performed by one or more computers and for synthesizing programs for generating images, the method comprising:
using a policy neural network to, at each of a sequence of time steps, select one or more actions for generating an image to be performed, wherein each action is a respective command to a graphics software configured to generate an image;
controlling the graphics software using the respective commands to cause the graphics software to generate a graphics output comprising an image;
processing the graphics output generated by the graphics software using a discriminator neural network to produce a discriminator output that indicates whether the graphics output is from a set of training data comprising images or was generated by the graphics software; and
training the policy neural network using a reinforcement learning procedure according to a reward signal determined from the discriminator output; and
training the discriminator neural network according to the discriminator output.
12 . The method of claim 11 wherein the graphics software is a non-differentiable graphics engine.
13 . The method of claim 11 wherein the policy neural network has an internal state, the method further comprising:
determining an estimated return from the internal state using a value function neural network, and
wherein training the policy neural network comprises updating parameters of the policy neural network using the estimated return.
14 . The method of claim 11 wherein training the policy neural network comprises updating parameters of the policy neural network using stored trajectories from an experience buffer that stores trajectories from graphics outputs generated by a plurality of actors.
15 . The method of claim 11 , the method further comprising storing the graphics output in a reply buffer, and wherein the discriminator neural network is configured to sample from the replay buffer to provide the discriminator output.
16 . 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 synthesizing programs for generating images, the operations comprising:
using a policy neural network to, at each of a sequence of time steps, select one or more actions for generating an image to be performed, wherein each action is a respective command to a graphics software configured to generate an image;
controlling the graphics software using the respective commands to cause the graphics software to generate a graphics output comprising an image;
processing the graphics output generated by the graphics software using a discriminator neural network to produce a discriminator output that indicates whether the graphics output is from a set of training data comprising images or was generated by the graphics software; and
training the policy neural network using a reinforcement learning procedure according to a reward signal determined from the discriminator output; and
training the discriminator neural network according to the discriminator output.
17 . The non-transitory computer storage media of claim 16 wherein the graphics software is a non-differentiable graphics engine.
18 . The non-transitory computer storage media of claim 16 wherein the policy neural network has an internal state, the operations further comprising:
determining an estimated return from the internal state using a value function neural network, and
wherein training the policy neural network comprises updating parameters of the policy neural network using the estimated return.
19 . The non-transitory computer storage media of claim 16 wherein training the policy neural network comprises updating parameters of the policy neural network using stored trajectories from an experience buffer that stores trajectories from graphics outputs generated by a plurality of actors.
20 . The non-transitory computer storage media of claim 16 , the operations further comprising storing the graphics output in a reply buffer, and wherein the discriminator neural network is configured to sample from the replay buffer to provide the discriminator output.