Meta-learning adaptive intrinsic reward weighting for curiosity-driven reinforcement learning

dc.contributor.advisorShock, Jonathan
dc.contributor.advisorSmit, Andries
dc.contributor.authorZiki, Batsirayi Mupamhi
dc.date.accessioned2026-07-01T08:29:54Z
dc.date.available2026-07-01T08:29:54Z
dc.date.issued2026
dc.date.updated2026-07-01T08:27:30Z
dc.description.abstractFor both organisms and artificial agents, exploration is essential to continue learning and avoid becoming trapped in suboptimal behaviours. Reinforcement learning (RL) agents can also face exploration challenges in environments with sparse feedback. Curiosity-driven exploration algorithms can help address these challenges by providing intrinsic rewards based on the novelty of situations an agent encounters. These intrinsic rewards are typically combined with extrinsic rewards using a weighted sum with the parameter λ. However, fine-tuning λ for each task across multiple environments can become computationally expensive. We propose a meta-learning approach for automatic tuning of λ using a recurrent neural network (RNN) that dynamically outputs λ values. We call this RNN the reward combiner. The reward combiner was trained using evolutionary strategies on XLand-MiniGrid environments, where feedback is sparse. The fitness function was the total extrinsic reward obtained during the training phase of an agent. We used BYOL-Explore, a curiosity-driven exploration algorithm, for intrinsic reward generation. The reward combiner takes normalised extrinsic and intrinsic rewards as input, along with actions that provide task-specific context for λ selection. Trained on Unlock and Empty-16x16 environments, the reward combiner generalises across different grid sizes of the same task, outperforming baselines when tested on DoorKey environments. It also generalises across different tasks when tested on UnlockPickUp, where the objective differs from the training environments. Our approach achieves higher extrinsic returns at the end of training than curiosity-driven baselines across all test environments. Despite being tested only within XLand-MiniGrid environments, our results indicate this approach has potential to eliminate costly hyperparameter sweeps when switching to new tasks with similar mechanics.
dc.identifier.apacitationZiki, B. M. (2026). <i>Meta-learning adaptive intrinsic reward weighting for curiosity-driven reinforcement learning</i>. (). University of Cape Town ,Faculty of Science ,Department of Mathematics and Applied Mathematics. Retrieved from http://hdl.handle.net/11427/43437en_ZA
dc.identifier.chicagocitationZiki, Batsirayi Mupamhi. <i>"Meta-learning adaptive intrinsic reward weighting for curiosity-driven reinforcement learning."</i> ., University of Cape Town ,Faculty of Science ,Department of Mathematics and Applied Mathematics, 2026. http://hdl.handle.net/11427/43437en_ZA
dc.identifier.citationZiki, B.M. 2026. Meta-learning adaptive intrinsic reward weighting for curiosity-driven reinforcement learning. . University of Cape Town ,Faculty of Science ,Department of Mathematics and Applied Mathematics. http://hdl.handle.net/11427/43437en_ZA
dc.identifier.ris TY - Thesis / Dissertation AU - Ziki, Batsirayi Mupamhi AB - For both organisms and artificial agents, exploration is essential to continue learning and avoid becoming trapped in suboptimal behaviours. Reinforcement learning (RL) agents can also face exploration challenges in environments with sparse feedback. Curiosity-driven exploration algorithms can help address these challenges by providing intrinsic rewards based on the novelty of situations an agent encounters. These intrinsic rewards are typically combined with extrinsic rewards using a weighted sum with the parameter λ. However, fine-tuning λ for each task across multiple environments can become computationally expensive. We propose a meta-learning approach for automatic tuning of λ using a recurrent neural network (RNN) that dynamically outputs λ values. We call this RNN the reward combiner. The reward combiner was trained using evolutionary strategies on XLand-MiniGrid environments, where feedback is sparse. The fitness function was the total extrinsic reward obtained during the training phase of an agent. We used BYOL-Explore, a curiosity-driven exploration algorithm, for intrinsic reward generation. The reward combiner takes normalised extrinsic and intrinsic rewards as input, along with actions that provide task-specific context for λ selection. Trained on Unlock and Empty-16x16 environments, the reward combiner generalises across different grid sizes of the same task, outperforming baselines when tested on DoorKey environments. It also generalises across different tasks when tested on UnlockPickUp, where the objective differs from the training environments. Our approach achieves higher extrinsic returns at the end of training than curiosity-driven baselines across all test environments. Despite being tested only within XLand-MiniGrid environments, our results indicate this approach has potential to eliminate costly hyperparameter sweeps when switching to new tasks with similar mechanics. DA - 2026 DB - OpenUCT DP - University of Cape Town KW - reinforcement learning KW - exploration algorithms LK - https://open.uct.ac.za PB - University of Cape Town PY - 2026 T1 - Meta-learning adaptive intrinsic reward weighting for curiosity-driven reinforcement learning TI - Meta-learning adaptive intrinsic reward weighting for curiosity-driven reinforcement learning UR - http://hdl.handle.net/11427/43437 ER - en_ZA
dc.identifier.urihttp://hdl.handle.net/11427/43437
dc.identifier.vancouvercitationZiki BM. Meta-learning adaptive intrinsic reward weighting for curiosity-driven reinforcement learning. []. University of Cape Town ,Faculty of Science ,Department of Mathematics and Applied Mathematics, 2026 [cited yyyy month dd]. Available from: http://hdl.handle.net/11427/43437en_ZA
dc.language.isoen
dc.language.rfc3066eng
dc.publisher.departmentDepartment of Mathematics and Applied Mathematics
dc.publisher.facultyFaculty of Science
dc.publisher.institutionUniversity of Cape Town
dc.subjectreinforcement learning
dc.subjectexploration algorithms
dc.titleMeta-learning adaptive intrinsic reward weighting for curiosity-driven reinforcement learning
dc.typeThesis / Dissertation
dc.type.qualificationlevelMasters
dc.type.qualificationlevelMSc
Files
Original bundle
Now showing 1 - 1 of 1
Loading...
Thumbnail Image
Name:
thesis_sci_2026_ziki batsirayi mupamhi.pdf
Size:
3.93 MB
Format:
Adobe Portable Document Format
Description:
License bundle
Now showing 1 - 1 of 1
Loading...
Thumbnail Image
Name:
license.txt
Size:
1.72 KB
Format:
Item-specific license agreed upon to submission
Description:
Collections