Reinforcement learning finds shorter circuits for preparing four-qubit states

[Highland Builder]

An AI agent trained with reinforcement learning has found efficient ways to build and untangle multi qubit states. Quantum Zeitgeist's Rusty Flint reports on work by researchers at the University of Maryland, EPFL and the Max Planck Institute for the Physics of Complex Systems. The team includes Norbert Linke and Marin Bukov

The trick is in what the agent sees. Instead of the full quantum state, which grows exponentially with the number of qubits, it only looks at two qubit reduced density matrices. That keeps the amount of information it needs growing polynomially, which makes larger systems manageable

One headline result is a general circuit that can prepare any four qubit state using at most five two qubit gates, equivalent to ten CNOT gates. The agent itself is a transformer model with an actor critic design that respects the symmetry of swapping qubits. It was tested on systems of up to 16 qubits

On four to six qubits, it beat random, greedy and Qiskit baseline methods. The team also ran it on a real trapped ion processor, where it successfully disentangled Bell pair, GHZ and W states. The code has been released publicly, which is welcome

Shorter circuits mean fewer chances for errors, which is crucial on today's noisy hardware. It is another example of AI helping to design quantum circuits, following the Claude work we discussed recently. Machine learning and quantum computing seem to be finding plenty of common ground

Will AI become the standard way to optimise quantum circuits? And how far can this approach scale beyond 16 qubits?

Reinforcement Learning Cuts Gate Count For 4-qubit Statesquantumzeitgeist.com