IAB Network Planning: A Deep Learning Framework for 6G Wireless Systems

Rapid growth in mobile data traffic and bandwidth-intensive applications
is driving unprecedented densification in 5th Generation Mobile Networks
(5G) and beyond. Integrated Access and Backhaul (IAB) operating in the
Millimeter Wave (mmWave) spectrum offers abundant bandwidth and flexi-
ble wireless backhaul, but network planning remains NP-hard and mmWave
links are vulnerable to environmental disruptions. This thesis presents a
comprehensive framework for intelligent IAB network planning through a
systematic evolution of optimisation and learning methods.
We first establish baselines using Mixed-Integer Linear Programming
(MILP) formulations and greedy heuristics, achieving near-optimal solutions
for networks with up to 30 nodes while exposing exponential computational
growth for larger deployments. To improve scalability, we develop Deep
Q-Network (DQN) variants with action elimination, reducing the required
number of deployed nodes by 12–15% compared with heuristic baselines.
Building on this, we propose discrete adaptations of Soft Actor-Critic (SAC)
enhanced with transfer learning, achieving a 20% reduction in deployed nodes
and reducing training time by 50% when adapting to new configurations.
Theoretical analysis establishes polynomial sample complexity O(|S||A|/ε2)
and preserves convergence guarantees under transfer learning.
Another contribution of our work is reformulating IAB planning as a
graph decision process and introducing a graph-centric policy based on Graph
Attention Network v2 (GATv2) with edge-conditioned attention. This en-
ables resilience-aware deployment that maintains 87.1% coverage under 30%
link failures (15.4% improvement over state-of-the-art) while reducing node
requirements by 26.7%. The framework achieves linear computational com-
plexity O(E · dh), making large scale deployments feasible. Finally, we dis-
cuss practical integration with Open Radio Access Network (O-RAN) via
the Non-RT RIC, supporting deployability in next-generation disaggregated
networks