Background

Reinforcement learning for network control faces strict requirements on sample efficiency, stability, and constraint satisfaction. Model-free methods often require prohibitive interaction time and struggle to generalize across traffic regimes and topologies. Model-based reinforcement learning addresses these limitations by explicitly learning a predictive world model of network dynamics and using this model for planning. Recent approaches such as latent-space world models and imagination-based rollouts have demonstrated that planning over learned dynamics can drastically reduce real-environment interactions. In parallel, planning-centric RL methods that combine learned models with search, such as those inspired by AlphaZero, have shown that coupling value and policy learning with lookahead planning yields robust and high-performing decision-makers. For networking problems with structured dynamics and well-defined constraints, model-based RL with planning is a particularly promising paradigm.

Problem Description

The objective of this student work is to investigate model-based reinforcement learning and planning for dynamic network control. A world model of the network dynamics is learned from interaction or offline data and subsequently used for planning-based decision making. The design is inspired by latent-dynamics models and imagination rollouts as well as search-based policy improvement. The work consists of the following tasks:

Acquired Knowledge and Skills

The student will gain a deep understanding of model-based reinforcement learning and planning under uncertainty. They will acquire practical experience with learning world models, latent-state representations, and planning-based policy optimization. The work further builds expertise in advanced RL architectures, sample-efficient learning, and dynamic network control, as well as strong skills in experimental design and evaluation of learning-based systems.