Data underlying the publication: Reinforcement learning in compilers for DQC

It becomes increasingly difficult to scale the number of qubits in a single quantum processor. In order to scale to more computational capability, distributed quantum computing (DQC) offers an alternative approach: connecting smaller modules into a network. In this architecture, the main challenge shifts from adding more qubits within one module to using qubits more efficiently across multiple modules. In particular, networking dynamics constrain this task as inter-module gates consume shared entangled states whose generation is typically much slower than local operations. These constraints make compilation for DQC a complex optimization problem, in which routing and scheduling decisions are tightly coupled over time. This sequential and resource-constrained structure makes reinforcement learning (RL) a promising approach for quantum circuit compilation.<br>In this work, we present a comprehensive numerical evaluation of the reinforcement learning approach proposed by Promponas et al. (2024) and compare it with our own agent, which incorporates a novel action-space formulation and effective action-masking strategy. Our results demonstrate that our agent achieves a reduction in the modeled execution time of up to 35\%. Finally, we carry out a critical assessment of both agents based on our results, lay out identified limitations and outline directions toward improving scalability and applicability of RL-based compilation for DQC.