Authors: Serena Ma, Chunwei Xia, and Zheng Wang (Univeristy of Leeds)
Abstract: Every circuit gate in a quantum program can add execution errors on noisy quantum hardware, making gate-count reduction important for reliable quantum computing. Rule-based optimizers use greedy search and miss improvements requiring long rewrite sequences. Learning-based methods are promising but constrained by incomplete circuit representations, unfocused exploration, and poor scalability. To address these limitations, we introduce GUIDERQ, a new reinforcement-learning optimizer with three key advances: combining local and global circuit features for richer representations, learned filtering to focus on the most promising candidate actions, and recursive partition-and-reassembly that scales to large circuits while exploring optimization opportunities across partitioning boundaries. Across representative quantum-circuit benchmarks with up to 27,883 gates and three industry-standard gate sets, GUIDERQ achieves geometric-mean gate-count reductions of 33.0%, 34.2%, and 25.1% on the IBM, Nam, and TDG gate sets — up to 34.1% more than Quarl, the state-of-the-art learning-based optimizer — with up to 3× faster search.
Presentation: file