Akademska digitalna zbirka SLovenije - logo
VSE knjižnice (vzajemna bibliografsko-kataložna baza podatkov COBIB.SI)
  • Qualitatively constrained control policy learning [Elektronski vir]
    Šoberl, Domen ; Žabkar, Jure, računalničar
    This paper introduces a novel approach that leverages qualitative reasoning to enhance reinforcement learning in physical domains. Traditional reinforcement learning methods often suffer from sample ... inefficiency and lack of explainability, especially in complex, physics-driven environments. Our proposed approach addresses these challenges by integrating qualitative induction and qualitative planning to learn a control strategy. Our method enables faster convergence of the learning process and yields an interpretable and physically plausible model of the environment. It employs a unique feedback loop mechanism that iteratively improves the qualitative model of the environment based on the observed outcomes of the executed plans, allowing continual refinement of the system’s understanding and actions. Through an extensive set of experiments, we demonstrate a superior performance of our method compared to a state-of-the-art deep reinforcement learning method.
    Vir: 36th International Workshop on Qualitative Reasoning : October 1st, 2023, Kraków [Elektronski vir] (Str. 1-8)
    Vrsta gradiva - prispevek na konferenci ; neleposlovje za odrasle
    Leto - 2023
    Jezik - angleški
    COBISS.SI-ID - 183909123

    Povezava(-e):

    https://staff.fnwi.uva.nl/b.bredeweg/QR2023/pdf/04Soberl.pdf

Vnos na polico