arXiv:2610.00945
TATVA: A Reinforcement Learning Framework for Quantum Circuit Synthesis
Om Bhamare, Aryan Jadhav, Manas Shinde, Prithvi Shinde
Abstract
One of the most important initial steps in quantum computing is high-fidelity quantum state preparation, because errors in it can affect the final computational result. Therefore, one of the major challenges is to design a quantum circuit that can achieve the desired target state accurately with a smaller number of gates. Traditionally, circuits are designed using predefined rules and mathematical decompositions, but this makes it difficult to design circuits for different and complex target states. This paper presents TATVA, a reinforcement learning (RL) system for synthesizing quantum circuits. TATVA views circuit synthesis as a sequential decision problem in which the agents select one gate at a time and calculate the fidelity after each applied gate, with the gate producing the highest fidelity selected for the circuit. The system introduces a parallel architecture that uses two RL agents, Deep Q-Network (DQN) and Proximal Policy Optimization (PPO), along with Qiskit's statevector simulator. The system can synthesize up to 5 qubits while achieving a fidelity of 0.999999 and producing compact circuits. After achieving the desired fidelity, it performs a post-hoc optimization step that reduces the circuit depth while preserving the achieved fidelity. The circuit is finalized using fidelity, success rate, gate count, and circuit depth. Target states that are not used during training are used to test whether the system has learned effectively and can generalize beyond the training states. This approach aims to enable automatic quantum circuit synthesis with high fidelity using reinforcement learning.
AI Summary
TATVA addresses automated quantum state preparation: finding gate sequences that approximate a supplied target state while limiting gate count and circuit depth. It formulates synthesis as a reinforcement-learning problem, starting from $|0\rangle^{\otimes n}$ and appending gates using DQN and PPO agents. Qiskit statevector simulation supplies the fidelity $F=|\langle\psi_{\mathrm{target}}|\psi_{\mathrm{current}}\rangle|^2$, and rewards combine fidelity, a gate-count penalty, and a success bonus. The two agents generate candidate circuits, which are selected according to fidelity and size. Post-processing removes redundant operations, merges rotations, and uses continuous L-BFGS-B parameter optimization. The reported simulations cover one to five qubits, including basis, superposition, Bell, GHZ, W, and Haar-random states. Reported pre-optimization fidelities are $0.999999$, rising to a numerical $1.0000000000$ after optimization. The results table gives gate-count reductions of $18\to4$, $38\to7$, $96\to47$, and $240\to114$ for two through five qubits. Five-qubit CNOT count decreases from 58 to 26 and depth from 214 to 106. Tests on held-out Haar-random states are described as producing candidates refinable by continuous optimization, but quantitative generalization rates and numerical comparisons against the stated classical baselines are not provided. The demonstrated scope is limited to small, noiseless statevector simulations. Scaling to larger systems, incorporating noise and hardware constraints, and execution on physical devices are left to future work.
Central claim may not hold: The reported table lacks sample counts, target-category breakdowns, and success rates needed to establish consistently successful synthesis across the stated evaluation ensemble. Continuous L-BFGS-B optimization is introduced in the results without sufficient methodological detail or ablations to establish the contribution of RL gate selection to the final circuit quality. Successful generalization of the learned policies is asserted without quantitative held-out results or evidence separating learned gate selection from subsequent continuous optimization. The results say agents selected CZ and SWAP even though neither gate belongs to the specified action set. The abstract's highest-immediate-fidelity gate-selection rule does not match the DQN and PPO action-selection procedures described in the body.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- 9 pages, 10 figures