Quarl: A Learning-Based Quantum Circuit Optimizer
summary
Quarl is a reinforcement-learning quantum circuit optimizer that replaces greedy search with a policy trained via PPO, decomposes millions of actions into a gate-selection decision plus a transformation-selection decision, and encodes local neighborhoods with a graph neural network. It trains online on a small set of representative circuits (order of ten) and uses policy-guided search and hierarchical advantage estimation. Evaluation on standard benchmarks reports up to 177% improvement over prior optimizers and Quarl learns nonlocal passes like rotation merging that previous tools treat as separate.
AI-written summary · about
chapters (11)
- 0:00Quarl optimizer intro
- 2:11Existing optimizer limits
- 6:04Why RL for optimization
- 7:29RL problem formulation
- 8:11Challenges in applying reinforcement learning
- 9:16Action space decomposition
- 10:20Gate state representation
- 12:08Quarl optimizer architecture
- 14:38Training with PPO
- 15:45Experimental results
- 16:28Q&A
related talks
21mQuantum Reinforcement Learning for Aerospace
31mReinforcement Learning with Quantum Algorithms: Simulating Nematode Behaviour
29mA Verified Optimizer for Quantum Circuits
19mReinforcement Learning for Optimizing Hybrid Quantum-Classical Workflows
19mEquality Saturation for Quantum Circuit Optimization
24mSynthesizing Quantum-Circuit Optimizers
22mHow to Synthesize Quantum-Circuit Optimizers
17mQuantum Applications with Classiq