Meta-Gradients in Non-Stationary Environments

Jelena Luketina, Sebastian Flennerhag, Yannick Schroecker, David Abel, Tom Zahavy, Satinder Singh
Proceedings of The 1st Conference on Lifelong Learning Agents, PMLR 199:886-901, 2022.

Abstract

Meta-gradient methods (Xu et al., 2018; Zahavy et al., 2020) offer a promising solution to the problem of hyperparameter selection and adaptation in non-stationary reinforcement learning problems. However, the properties of meta-gradients in such environments have not been systematically studied. In this work, we bring new clarity to meta-gradients in non-stationary environments. Concretely, we ask: (i) how much information should be given to the learned optimizers, so as to enable faster adaptation and generalization over a lifetime, (ii) what meta-optimizer functions are learned in this process, and (iii) whether meta-gradient methods provide a bigger advantage in highly non-stationary environments. To study the effect of information provided to the meta-optimizer, as in recent works (Flennerhaget al., 2021; Almeida et al., 2021), we replace the tuned meta-parameters of fixed update rules with learned meta-parameter functions of selected context features. The context features carry information about agent performance and changes in the environment and hence can inform learned meta-parameter schedules. We find that adding more contextual information is generally beneficial, leading to faster adaptation of meta-parameter values and increased performance over a lifetime. We support these results with a qualitative analysis of resulting meta parameter schedules and learned functions of context features. Lastly, we find that without context, meta-gradients do not provide a consistent advantage over the baseline in highly non-stationary environments. Our findings suggest that contextualizing meta-gradients can play a pivotal role in extracting high performance from meta-gradients in non-stationary settings.

Cite this Paper

BibTeX
@InProceedings{pmlr-v199-luketina22a, title = {Meta-Gradients in Non-Stationary Environments}, author = {Luketina, Jelena and Flennerhag, Sebastian and Schroecker, Yannick and Abel, David and Zahavy, Tom and Singh, Satinder}, booktitle = {Proceedings of The 1st Conference on Lifelong Learning Agents}, pages = {886--901}, year = {2022}, editor = {Chandar, Sarath and Pascanu, Razvan and Precup, Doina}, volume = {199}, series = {Proceedings of Machine Learning Research}, month = {22--24 Aug}, publisher = {PMLR}, pdf = {https://proceedings.mlr.press/v199/luketina22a/luketina22a.pdf}, url = {https://proceedings.mlr.press/v199/luketina22a.html}, abstract = {Meta-gradient methods (Xu et al., 2018; Zahavy et al., 2020) offer a promising solution to the problem of hyperparameter selection and adaptation in non-stationary reinforcement learning problems. However, the properties of meta-gradients in such environments have not been systematically studied. In this work, we bring new clarity to meta-gradients in non-stationary environments. Concretely, we ask: (i) how much information should be given to the learned optimizers, so as to enable faster adaptation and generalization over a lifetime, (ii) what meta-optimizer functions are learned in this process, and (iii) whether meta-gradient methods provide a bigger advantage in highly non-stationary environments. To study the effect of information provided to the meta-optimizer, as in recent works (Flennerhaget al., 2021; Almeida et al., 2021), we replace the tuned meta-parameters of fixed update rules with learned meta-parameter functions of selected context features. The context features carry information about agent performance and changes in the environment and hence can inform learned meta-parameter schedules. We find that adding more contextual information is generally beneficial, leading to faster adaptation of meta-parameter values and increased performance over a lifetime. We support these results with a qualitative analysis of resulting meta parameter schedules and learned functions of context features. Lastly, we find that without context, meta-gradients do not provide a consistent advantage over the baseline in highly non-stationary environments. Our findings suggest that contextualizing meta-gradients can play a pivotal role in extracting high performance from meta-gradients in non-stationary settings.} }
Endnote
%0 Conference Paper %T Meta-Gradients in Non-Stationary Environments %A Jelena Luketina %A Sebastian Flennerhag %A Yannick Schroecker %A David Abel %A Tom Zahavy %A Satinder Singh %B Proceedings of The 1st Conference on Lifelong Learning Agents %C Proceedings of Machine Learning Research %D 2022 %E Sarath Chandar %E Razvan Pascanu %E Doina Precup %F pmlr-v199-luketina22a %I PMLR %P 886--901 %U https://proceedings.mlr.press/v199/luketina22a.html %V 199 %X Meta-gradient methods (Xu et al., 2018; Zahavy et al., 2020) offer a promising solution to the problem of hyperparameter selection and adaptation in non-stationary reinforcement learning problems. However, the properties of meta-gradients in such environments have not been systematically studied. In this work, we bring new clarity to meta-gradients in non-stationary environments. Concretely, we ask: (i) how much information should be given to the learned optimizers, so as to enable faster adaptation and generalization over a lifetime, (ii) what meta-optimizer functions are learned in this process, and (iii) whether meta-gradient methods provide a bigger advantage in highly non-stationary environments. To study the effect of information provided to the meta-optimizer, as in recent works (Flennerhaget al., 2021; Almeida et al., 2021), we replace the tuned meta-parameters of fixed update rules with learned meta-parameter functions of selected context features. The context features carry information about agent performance and changes in the environment and hence can inform learned meta-parameter schedules. We find that adding more contextual information is generally beneficial, leading to faster adaptation of meta-parameter values and increased performance over a lifetime. We support these results with a qualitative analysis of resulting meta parameter schedules and learned functions of context features. Lastly, we find that without context, meta-gradients do not provide a consistent advantage over the baseline in highly non-stationary environments. Our findings suggest that contextualizing meta-gradients can play a pivotal role in extracting high performance from meta-gradients in non-stationary settings.
APA
Luketina, J., Flennerhag, S., Schroecker, Y., Abel, D., Zahavy, T. & Singh, S.. (2022). Meta-Gradients in Non-Stationary Environments. Proceedings of The 1st Conference on Lifelong Learning Agents, in Proceedings of Machine Learning Research 199:886-901 Available from https://proceedings.mlr.press/v199/luketina22a.html.

Related Material