Recurrent World Models for RAN Optimization

Sammanfattning

Modern 5G base stations choose a Modulation and Coding Scheme (MCS) for every transmission slot. Production schedulers do this with a static lookup table indexed by a channel-quality indicator, which does not adapt to the current channel realisation. This thesis asks whether a small data-driven agent, trained on a learned latent model of the channel, can do better. The architecture adapts the Recurrent World Models design of Ha and Schmidhuber to 5G channel data. A convolutional autoencoder compresses the per-subcarrier Signal-to-Interference-plus-Noise Ratio (SINR) vector into a sixteen-dimensional latent. A mixture-density recurrent network reads the latent and the previous action and summarises history into a hidden state. A single linear controller reads both the latent and the hidden state and emits the MCS index. The controller is trained by covariance-matrix-adaptation evolution strategy on a 5G Urban Macro and Urban Micro simulator dataset. Across all 96 rollouts of 200 slots each in the simulator dataset, the trained controller outperforms every one of the 22 fixed-MCS policies and uniform-random MCS selection in cumulative normalised throughput, with the learned controller's curve sitting above the upper envelope of the fixed-MCS curves throughout. The decomposition has a secondary structural property worth noting: because the autoencoder and the recurrent model are trained with unsupervised losses, reward-driven learning is confined to a small linear controller, which is expected to keep the approach light on reward-bearing simulator interactions. A quantitative comparison against an end-to-end deep reinforcement learning baseline on matched data is not made here. The thesis does not claim to beat the current state of the art; direct comparisons against classical outer-loop link adaptation and against published latent-bandit schemes are left to future work.