Commit History
30 days ago
-
Add FUMD-AI handover forecaster (combined run)
Verified signatureInitial release of the multi-step serving-cell forecaster: given 6 s of a vehicle's joint SUMO mobility and OMNeT++/Simu5G radio history, it predicts the serving base station at each of the next 1-7 s. Architecture: a 2x bidirectional-LSTM(128) encoder feeding a per-step Bahdanau-attention(128) decoder, autoregressive over the horizon via an Embedding(10, 128) of the previous step's argmax. ~1.14 M parameters, 7 softmax heads over 10 base stations, saved as a Keras 2.14 .keras archive. Weights are the pipeline_run_combined output of the FUMD-AI training & postprocessing workflow v0.1.4, trained on the four combined _1 datasets (900/1000/1200/1400 vehicles) from one simulated urban scenario (Alicante, 10 cells): 1,085,658 training windows and 271,759 validation windows, split by vehicle rather than by window, so no validation window shares a source row with training. Adam 1e-3 with gradient clipping, batch 64; early stopping restored the epoch-6 weights after 9 epochs. Held-out valid…b020cf2
