Knowledge Transfer in Deep Reinforcement Learning via an RL-Specific GAN-Based Correspondence Function

0602819 20250317083328.420241216235959.9 85210284955 001370660700039 10.1109/ACCESS.2024.3497589 Knowledge Transfer in Deep Reinforcement Learning via an RL-Specific GAN-Based Correspondence Function 15 s. E cav_un_epca*04610362169-3536IEEE Access121 (2024)177204-177218Institute of Electrical and Electronics Engineers Deep learning Markov decision process reinforcement learning transfer learning knowledge transfer cav_un_auth*0333672 Ruman Marko UTIA-B Adaptivní systémy Department of Adaptive Systems AS AS SK K Ústav teorie informace a automatizace AV ČR, v. v. i. cav_un_auth*0101092 Guy Tatiana Valentine UTIA-B Adaptivní systémy Department of Adaptive Systems AS AS Department of Adaptive Systems Ústav teorie informace a automatizace AV ČR, v. v. i. https://library.utia.cas.cz/separaty/2024/AS/guy-0602819.pdf https://ieeexplore.ieee.org/document/10752398 CA21169 EU-COST XE cav_un_auth*0452289 Deep reinforcement learning has demonstrated superhuman performance in complex decision-making tasks, but it struggles with generalization and knowledge reuse—key aspects of true intelligence. This article introduces a novel approach that modifies Cycle Generative Adversarial Networks specifically for reinforcement learning, enabling effective one-to-one knowledge transfer between two tasks. Our method enhances the loss function with two new components: model loss, which captures dynamic relationships between source and target tasks, and Q-loss, which identifies states significantly influencing the target decision policy. Tested on the 2-D Atari game Pong, our method achieved 100% knowledge transfer in identical tasks and either 100% knowledge transfer or a 30% reduction in training time for a rotated task, depending on the network architecture. In contrast, using standard Generative Adversarial Networks or Cycle Generative Adversarial Networks led to worse performance than training from scratch in the majority of cases. The results demonstrate that the proposed method ensured enhanced knowledge generalization in deep reinforcement learning. WOS IN 10000 10200 10201 2025 2 2 4 R hod 4 4rh 4 20250310143107.1 20250310150006.2 RVO:67985556 https://hdl.handle.net/11104/0360153 cav_un_auth*0478849 Provozně ekonomická fakulta, Česká zemědělská univerzita v Praze PEF CZU CZ S A TELECOMMUNICATIONS|ENGINEERING.ELECTRICAL&ELECTRONIC|COMPUTERSCIENCE.INFORMATIONSYSTEMS 3.9 0.8 4.1 0.3457 0.67 294150 290 0.849 50.78 5.31 185847 Q1 3.200 13193 Q2 62.3 0.83 Q2 64.8 2024 Soubory v repozitáři: guy-0602819.pdf 85210284955 SCOPUS PUBMED 001370660700039 WOS cav_un_epca*0461036 IEEE Access Roč. 12 č. 1 2024 177204 177218 2169-3536 2169-3536 Institute of Electrical and Electronics Engineers