Boosting deep reinforcement learning performance by combining off-line data and simulators
Example implementations described herein involve an approach to address an imperfect simulator challenge using off-line data plus reward modification. The proposed solution is robust to simulator error, and therefore, it requires less maintenance in keeping the simulators updated. Even when the simulators are accurate, it is costly to keep them accurate over time. Moreover, compared to other robust reinforcement learning algorithms, the proposed approach does not assume the distribution of uncertainties in the simulator are known. Less complexity leads to fewer potential errors as well as lower computational cost during the training. Finally, the proposed approach has better performance compared to the state-of-the-art methods (higher overall cumulative rewards).
1 . A method for training of a reinforcement learning policy, comprising:
executing in parallel an ensemble of models and a simulator to predict a next state, the ensemble of models trained from using off-line data;
for predictions of the next state from the execution of the ensemble of models being in disagreement when a standard deviation of outputs from the ensemble of models exceeds a threshold:
using the next state predicted from the execution of the simulator as a sample for training the reinforcement learning policy; and
imposing a penalty on a reward for the training the reinforcement learning policy based on a variance of the predictions of the next state from the ensemble of models;
for the predictions of the next state from the execution of the ensemble of models being in agreement when the standard deviation of outputs from the ensemble of models is within the threshold:
using weighted predictions of the next state from the execution of the ensemble of models as the sample for training the reinforcement learning policy; and
adjusting the reward on the ensemble of models based on the variance of the predictions of the ensemble of models and a difference between the predictions of the next state from the execution of the ensemble of models and the next state predicted from the execution of the simulator.
2 . The method of claim 1 , wherein the difference between the predictions of the next state from the execution of the ensemble of models and the next state predicted from the execution of the simulator is a difference between a weighted average of the predictions of the next state from the execution of the ensemble of models and the next state predicted from the execution of the simulator.
3 . The method of claim 1 , wherein the off-line data comprises off-line images, wherein the method further comprises training an encoder to translate the off-line images and output images from the simulator into a common domain.
4 . The method of claim 3 , wherein the off-line data comprises off-line images, wherein the ensemble of models is trained from encoded images in the common domain generated from the encoder as executed on the off-line images, and wherein the ensemble of models is configured to predict the next state as an image representation in the common domain;
wherein the difference between the predictions of the next state from the execution of the ensemble of models and the next state predicted from the execution of the simulator is determined based on executing the encoder on the output images from the simulator to translate the output images to the common domain and determining the difference between the translated output images in the common domain and the image representation.
5 . The method of claim 1 , when the uncertainty of the simulator is known, wherein the using the weighted predictions of the next state from the execution of the ensemble of models as the sample for training the reinforcement learning algorithm comprises:
using the weighted predictions of the next state from the execution of the ensemble of models and the next state predicted from the execution of the simulator as samples for training the reinforcement learning algorithm based on uncertainty of the simulator and uncertainty of the ensemble of models.
6 . The method of claim 1 , wherein the predictions of the next state from the execution of the ensemble of models are in disagreement when the variance exceeds a threshold.
7 . The method of claim 1 , wherein the predictions of the next state from the execution of the ensemble of models are in agreement when the variance is within a threshold.
8 . The method of claim 1 , further comprising:
deploying the reinforcement learning policy; and
periodically training the reinforcement learning policy through the executing of the ensemble of models and the simulator based on accumulation of new off-line data.
9 . The method of claim 1 , further comprising:
initializing a policy of the reinforcement learning policy;
wherein the executing the ensemble of models and the simulator to predict the next state is conducted according to the policy;
wherein the sample for training the reinforcement learning policy is accumulated into a sample set from which a batch of samples from the sample set are used to train the reinforcement learning algorithm.
10 . A non-transitory computer readable medium, storing instructions for training of a reinforcement learning policy, the instructions comprising:
executing in parallel an ensemble of models and a simulator to predict a next state, the ensemble of models trained from using off-line data;
for predictions of the next state from the execution of the ensemble of models being in disagreement when a standard deviation of outputs from the ensemble of models exceeds a threshold:
using the next state predicted from the execution of the simulator as a sample for training the reinforcement learning policy; and
imposing a penalty on a reward for the training the reinforcement learning policy based on a variance of the predictions of the next state from the ensemble of models;
for the predictions of the next state from the execution of the ensemble of models being in agreement when the standard deviation of outputs from the ensemble of models is within the threshold:
using weighted predictions of the next state from the execution of the ensemble of models as the sample for training the reinforcement learning policy; and
adjusting the reward on the ensemble of models based on the variance of the predictions of the ensemble of models and a difference between the predictions of the next state from the execution of the ensemble of models and the next state predicted from the execution of the simulator.
11 . The non-transitory computer readable medium of claim 10 , wherein the difference between the predictions of the next state from the execution of the ensemble of models and the next state predicted from the execution of the simulator is a difference between a weighted average of the predictions of the next state from the execution of the ensemble of models and the next state predicted from the execution of the simulator.
12 . The non-transitory computer readable medium of claim 10 , wherein the off-line data comprises off-line images, wherein the instructions further comprises training an encoder to translate the off-line images and output images from the simulator into a common domain.
13 . The non-transitory computer readable medium of claim 12 , wherein the off-line data comprises off-line images, wherein the ensemble of models is trained from encoded images in the common domain generated from the encoder as executed on the off-line images, and wherein the ensemble of models is configured to predict the next state as an image representation in the common domain;
wherein the difference between the predictions of the next state from the execution of the ensemble of models and the next state predicted from the execution of the simulator is determined based on executing the encoder on the output images from the simulator to translate the output images to the common domain and determining the difference between the translated output images in the common domain and the image representation.
14 . The non-transitory computer readable medium of claim 10 , when the uncertainty of the simulator is known, wherein the using the weighted predictions of the next state from the execution of the ensemble of models as the sample for training the reinforcement learning algorithm comprises:
using the weighted predictions of the next state from the execution of the ensemble of models and the next state predicted from the execution of the simulator as samples for training the reinforcement learning algorithm based on uncertainty of the simulator and uncertainty of the ensemble of models.
15 . The non-transitory computer readable medium of claim 10 , wherein the predictions of the next state from the execution of the ensemble of models are in disagreement when the variance exceeds a threshold.
16 . The non-transitory computer readable medium of claim 10 , wherein the predictions of the next state from the execution of the ensemble of models are in agreement when the variance is within a threshold.
17 . The non-transitory computer readable medium of claim 10 , further comprising:
deploying the reinforcement learning policy; and
periodically training the reinforcement learning policy through the executing of the ensemble of models and the simulator based on accumulation of new off-line data.
18 . The non-transitory computer readable medium of claim 10 , further comprising:
initializing a policy of the reinforcement learning policy;
wherein the executing the ensemble of models and the simulator to predict the next state is conducted according to the policy;
wherein the sample for training the reinforcement learning policy is accumulated into a sample set from which a batch of samples from the sample set are used to train the reinforcement learning algorithm.
19 . An apparatus, configured to train a reinforcement learning policy, the apparatus comprising:
a processor, configured to:
execute in parallel an ensemble of models and a simulator to predict a next state, the ensemble of models trained from using off-line data;
for predictions of the next state from the execution of the ensemble of models being in disagreement when a standard deviation of outputs from the ensemble of models exceeds a threshold:
use the next state predicted from the execution of the simulator as a sample for training the reinforcement learning policy; and
impose a penalty on the reward for the training the reinforcement learning algorithm based on a variance of the predictions of the next state from the ensemble of models;
for the predictions of the next state from the execution of the ensemble of models being in agreement when the standard deviation of outputs from the ensemble of models is within the threshold:
use weighted predictions of the next state from the execution of the ensemble of models as the sample for training the reinforcement learning policy; and
adjust the reward on the ensemble of models based on the variance of the predictions of the ensemble of models and a difference between the predictions of the next state from the execution of the ensemble of models and the next state predicted from the execution of the simulator.