NASA’s new dark energy space telescope can also detect killer asteroids
technologyreview.com·2h ago
AI Research
arXiv CS.AI·May 25·8 min read
Classical State Preparation for Variational Quantum Algorithms via Reinforcement Learning
TL;DR
Variational Quantum Algorithms (VQAs)—hybrid classical-quantum methods for optimization—suffer from barren plateaus (flat loss landscapes) and poor initialization. CRiSP, a reinforcement learning framework, uses a Transformer-based policy with Monte Carlo Tree Search to automatically select Clifford gates (classically simulable quantum operations) that warm-start VQAs.
✦ Why It Matters
Engineers can use CRiSP to automatically initialize quantum circuits with better convergence, reducing quantum hardware time and improving practical quantum algorithm performance.
Key Takeaways
- CRiSP automates Clifford gate selection for VQA initialization using reinforcement learning and tree search.
- Achieves 3.17Ă— mean and up to 45Ă— maximum improvement over prior Clifford initialization methods.
- Runs entirely classically in polynomial time; no quantum hardware needed for state preparation.
- Curriculum learning strategy enables scaling to deep circuits with 1,370+ parameters on 22+ qubits.
Full Summary
Variational Quantum Algorithms combine classical optimization with quantum circuits to solve problems on near-term quantum hardware, but struggle with barren plateaus—regions where gradients vanish—and poor convergence. Warm-starting VQAs with good initial quantum states can accelerate training, but existing heuristic methods fail to scale across the exponentially large space of possible initializations.
CRiSP (Clifford Reinforcement Learning agent for State Preparation) formulates gate selection as a sequential decision problem, using a Transformer neural network trained via self-play to guide Monte Carlo Tree Search in choosing which Clifford gates to prepend before fixed parameterized rotations. The approach runs entirely in polynomial time using classical stabilizer simulation, avoiding quantum hardware during initialization.
Curriculum learning progressively expands search depth to handle deep circuits. Evaluated on Quantum Approximate Optimization Algorithm (QAOA) benchmarks with up to 22 qubits and 1,370 parameters, CRiSP outperformed state-of-the-art Clifford initialization by 3.17Ă— in mean energy accuracy and 2.44Ă— in best-achieved accuracy, with maximum improvements reaching 45Ă— and 16Ă— respectively.