An intelligent reinforcement learning framework for sequential decision-making
Applied Soft Computing, cilt.202, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 202
- Basım Tarihi: 2026
- Doi Numarası: 10.1016/j.asoc.2026.115800
- Dergi Adı: Applied Soft Computing
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Applied Science & Technology Source, Compendex, INSPEC
- Anahtar Kelimeler: Deep reinforcement learning, LSTM, Sequential decision-making, Transformer, Uncertainty
- Bursa Uludağ Üniversitesi Adresli: Evet
Özet
Reliable decision-making under uncertainty remains a core challenge in deep reinforcement learning. A unified evaluation framework is presented to probe decision-making under epistemic and aleatoric uncertainty across four observability regimes using a standardized protocol. Representative model-free algorithms (policy- and value-based) are assessed with long short-term memory (LSTM) and transformer function approximators to isolate how architecture and uncertainty type interact. An enhanced deep Q-learning variant is also proposed as a baseline. Across uncertain settings, LSTM backbones often match or exceed transformer performance; noisy inputs are typically easier than limited-feedback settings, while belief-state outcomes vary by task and backbone. These findings suggest that robustness depends jointly on the learning algorithm and function approximator; within this framework, the included E-DQL baseline attains top or tied success rates in a majority of settings, with environment- and backbone-specific exceptions. Moreover, the side-by-side treatment of uncertainties, together with the systematic comparison of LSTM and transformer backbones across these regimes, is provided as a resource expected to aid future studies.