The thinking behind the RA-NN.
These papers define the intelligence model behind the RA-NN — how the radio environment is represented, and how control decisions are derived from that representation. The CU-IP is where this intelligence lives, as a native function inside the gNB.
Applying the RA-NN Formulation to PCI Optimization
Michael Chiaramonte
PCI optimization has been misformulated as graph coloring — the wrong graph, the wrong objective, and the wrong evaluation. This paper applies the RA-NN formulation to PCI planning, constructing a sparse interaction graph from UE measurement reports and targeting confusion (the actual operational problem) instead of collisions. The resolution procedure converges monotonically, executes in under 0.1 seconds on a 1,500-cell network, and changes only the cells that need to change.
Read the paper (PDF)
A Neural Systems Formulation for AI-Native Radio Networks
Michael Chiaramonte
This paper introduces the RA-NN — a neural systems formulation that treats the radio access network itself as a learnable system. Rather than applying AI models on top of legacy RAN abstractions, the RA-NN defines a new architectural foundation where the network learns its own structure from per-UE observations, builds a shared latent substrate, and derives control policies directly from that substrate.
Read the paper (PDF)