Browsing by Author "Wu, Xinyang"

Filter results by typing the first few letters
Now showing 1 - 1 of 1
  • Thumbnail Image
    ItemOpen Access
    Data-efficient reinforcement learning with Bayesian neural networks
    (Stuttgart : Fraunhofer-Institut für Produktionstechnik und Automatisierung IPA, 2025) Wu, Xinyang; Huber, Marco (Prof. Dr.-Ing. habil.)
    Artificial Intelligence (AI) and Machine Learning (ML) have propelled significant advancements across numerous domains, with deep Reinforcement Learning (RL) emerging as a critical solution for complex control tasks. While traditional Neural Networks (NNs) enhance performance and learning capacity, they often exhibit overconfidence and lack uncertainty information in their predictions, which can compromise optimal decision-making in stochastic environments. This thesis elucidates the potential of Bayesian Neural Networks (BNNs), which reconcile the predictive capacities of NNs with the probabilistic rigor of Bayesian inference, offering a robust paradigm for uncertainty quantification. An innovative approach is introduced, employing the Kalman filter, a powerful tool for state estimation in dynamic systems, to enable efficient online learning of BNNs. The effectiveness and efficiency of this approach are validated on standard ML datasets. Beyond providing a theoretical exposition of BNNs, the thesis pioneers the integration of BNNs within both model-free and model-based RL frameworks. The objective is to utilize the uncertainty quantification capabilities of BNNs to improve learning efficiency and safety performance of RL algorithms, over- coming challenges associated with overconfidence and uncertain predictions. The practical efficacy of the proposed methodologies is validated through experiments on classic control problems and complex robotic tasks. The empirical results underscore significant improvements in learning efficiency and safety performance, proving the theoretical merits of integrating BNNs with RL. In conclusion, this thesis offers an in-depth exploration into the fusion of BNNs and RL, presenting innovative methodologies that incorporate uncertainty information into RL paradigms. The insights and methodologies proposed serve as a springboard for future research, moving us closer to the realization of RL’s full potential in real-world applications.