Published in University of Antwerp, Faculty of Science, 2025
Manuscript of my PhD thesis. Study of unsupervised skill discovery and task transfer in reinforcement learning: learning behaviors and representations without reward to quickly maximize any downstream task reward.
Published in Transactions on Machine Learning Research, 2025
We introduce a novel method for unsupervised zero-shot RL based on the partitioning of the state space into clusters of temporal proximity. We employ the Successor Feature framework, which allows us to train a general policy and visualize the expected optimal trajectories. We achieve state-of-the-art zero-shot performance for SF-based methods on MuJoCo tasks.
Published in Adaptive Learning Agents Workshop of AAMAS 2023; Neural Computing and Applications, 2025
We study unsupervised skill discovery, and in particular the usage of learned skills during transfer. We propose to enhance the General Policy Improvement Theorem with a k-step forward search and show that it is possible to plan at the skill level with this approach and framework.
Published in BNAIC/BeNeLearn 2022, 2022
Successor Features propose to linearly decompose the reward function through a basis of rewards or features, but leave the exact features undefined. We study different possible types of features and their impact on learning and task representation.
Published in Deep Reinforcement Learning Workshop NeurIPS 2022, 2022
We extend Explore Options for Deep Reinforcement Learning, adapting the algorithms for neural-network-based function approximation through the off-policy training of multi-headed neural networks. We provide a clear framework on the different choices of combination of intrinsic and extrinsic rewards. We evaluate the methods on challenging Atari exploration games.
Published in 4th Lifelong Machine Learning Workshop at ICML 2020, 2020
We study intrinsic motivation for exploration in RL and propose an aternative to the wide-spread approach of a weighted sum of rewards. We show that decoupling exploration and exploitation through different agents enables to scale to multiple intrinsic rewards, ignore harmful signals and improve task transfer. We propose to switch between agents through an Explore Option (an additional to call the Explorer agent), leading to exploration-focused Hierarchical RL.