Multi-agent reinforcement learning with diffusion models
Πολυπρακτορική ενισχυτική μάθηση με μοντέλα διάχυσης
View/
Keywords
Deep Learning ; Reinforcement Learning ; Multi-agent Systems ; Diffusion ModelsAbstract
Diffusion models have been increasingly applied to Reinforcement Learning (RL) in order to deal with complex decision-making tasks. However, their effectiveness in learning multi-agent policies have not been thoroughly studied in the literature. This thesis explores how these models can enhance Multi-Agent RL (MARL) techniques in complex multi-agent environments under the celebrated CTDE schema. We present a MARL method, dubbed Q-Diffuser, which aims at inferring imaginative communication messages among agents, and further using meaningful inferred information to enhance the estimation of the Q-value function building upon the most premier MARL algorithm, called QMIX. The approach leverages a wide array of state-of-the-art techniques, including Denoising Diffusion Probabilistic Models (DDPM), transformer architectures, and the individual-global-max (IGM) property. Experimentally, we evaluate Q-Diffuser on the widely used StarCraft Multi-Agent Challenge (SMAC) benchmark and demonstrate superior performance over vanilla QMIX on a diverse set of challenging tasks, including Hard and Super-Hard maps.
Postgraduate Studies Programme
Τεχνητή Νοημοσύνη - Artificial IntelligenceDepartment
Σχολή Τεχνολογιών Πληροφορικής και Επικοινωνιών. Τμήμα Ψηφιακών ΣυστημάτωνCorporate Department
National Center of Scientific Research "Demokritos"Number of pages
85Language
EnglishCollections
Except where otherwise noted, this item's license is described as
Αναφορά Δημιουργού-Μη Εμπορική Χρήση-Όχι Παράγωγα Έργα 3.0 Ελλάδα
Αναφορά Δημιουργού-Μη Εμπορική Χρήση-Όχι Παράγωγα Έργα 3.0 Ελλάδα