%0 Journal Article %1 c9671066ca224d36b8f597fab5800076 %A Benjamins, Carolin %A Eimer, Theresa %A Schubert, Frederik Günter %A Mohan, Aditya %A Döhler, Sebastian %A Biedenkapp, André %A Rosenhahn, Bodo %A Hutter, Frank %A Lindauer, Marius %D 2023 %J Transactions on Machine Learning Research %K #sys:relevantfor:l3s automl myown %R 10.48550/arXiv.2202.04500 %T Contextualize Me – The Case for Context in Reinforcement Learning %X While Reinforcement Learning ( RL) has made great strides towards solving increasinglycomplicated problems, many algorithms are still brittle to even slight environmental changes.Contextual Reinforcement Learning (cRL) provides a framework to model such changes ina principled manner, thereby enabling flexible, precise and interpretable task specificationand generation. Our goal is to show how the framework of cRL contributes to improvingzero-shot generalization in RL through meaningful benchmarks and structured reasoningabout generalization tasks. We confirm the insight that optimal behavior in cRL requirescontext information, as in other related areas of partial observability. To empirically validatethis in the cRL framework, we provide various context-extended versions of common RLenvironments. They are part of the first benchmark library, CARL, designed for generalizationbased on cRL extensions of popular benchmarks, which we propose as a testbed to furtherstudy general agents. We show that in the contextual setting, even simple RL environmentsbecome challenging - and that naive solutions are not enough to generalize across complexcontext spaces.

@article{c9671066ca224d36b8f597fab5800076, abstract = { While Reinforcement Learning ( RL) has made great strides towards solving increasinglycomplicated problems, many algorithms are still brittle to even slight environmental changes.Contextual Reinforcement Learning (cRL) provides a framework to model such changes ina principled manner, thereby enabling flexible, precise and interpretable task specificationand generation. Our goal is to show how the framework of cRL contributes to improvingzero-shot generalization in RL through meaningful benchmarks and structured reasoningabout generalization tasks. We confirm the insight that optimal behavior in cRL requirescontext information, as in other related areas of partial observability. To empirically validatethis in the cRL framework, we provide various context-extended versions of common RLenvironments. They are part of the first benchmark library, CARL, designed for generalizationbased on cRL extensions of popular benchmarks, which we propose as a testbed to furtherstudy general agents. We show that in the contextual setting, even simple RL environmentsbecome challenging - and that naive solutions are not enough to generalize across complexcontext spaces.}, added-at = {2024-02-14T12:15:18.000+0100}, author = {Benjamins, Carolin and Eimer, Theresa and Schubert, {Frederik G{\"u}nter} and Mohan, Aditya and D{\"o}hler, Sebastian and Biedenkapp, Andr{\'e} and Rosenhahn, Bodo and Hutter, Frank and Lindauer, Marius}, biburl = {https://www.bibsonomy.org/bibtex/2ca5bf41777b83e3118bb63b8a9f12089/ail3s}, doi = {10.48550/arXiv.2202.04500}, interhash = {e0c5bde4915c039bf30a53c96114dcd7}, intrahash = {ca5bf41777b83e3118bb63b8a9f12089}, issn = {2835-8856}, journal = {Transactions on Machine Learning Research}, keywords = {#sys:relevantfor:l3s automl myown}, language = {English}, timestamp = {2024-02-27T12:16:27.000+0100}, title = {Contextualize Me – The Case for Context in Reinforcement Learning}, year = 2023 }