Gate model quantum computers require specialized software to translate quantum instructions into low-level instructions, i.e. compilers. Quantum compilers based on reinforcement learning techniques will provide a minimum number of instructions in a minimum execution time, which could be exploited for real-time quantum compiling and to minimize quantum noise.

Patent Status

SUBMITTED

Priority Number

102021000006179

Priority Date

16/03/2021

License

ITALY

Problem

The architecture of gate model quantum computers requires software computation layers dedicated to the compilation of high-level quantum algorithms in low-level quantum logic gate circuits. Due to quantum noise and the fragility of quantum properties that characterize computation, efficient compilation strategies are required to build optimized circuits quickly, reduce noise and optimize quantum resources. Traditional approaches require long computation times, not suitable during quantum computation. Using a quantum compiler based on artificial intelligence trained through multiple iterations of reinforcement learning allows creating single-qubit operations in a minimum computational time.

Current Technology Limitations

There are several quantum compiling strategies and algorithms with different performances. The performances can be measured and compared using three metrics: the number of logic gates used to build the circuit, the execution time, and the precompile time. Regardless of the quantum compiler employed, these metrics cannot simultaneously scale optimally. Compilers based on the Solovay-Kitaev theorem provide quantum circuits by balancing these three metrics, although several formulations and implementations have different performances. However, these strategies require specific low-level quantum logic gate sets.On the other hand, other compilation strategies such as KAK decomposition exhibit better performance but are specific to low-qubit circuit compilations. Recently, compilation strategies based on machine learning and artificial intelligence techniques have been proposed. While these techniques promise optimized circuits, they generally exhibit high execution times due to the limited number of precompilation steps they can employ.

Our Technology and Solution

The quantum compilation method allows compiling high-level quantum algorithms into low-level quantum logic gate circuits with a single learning procedure by using artificial intelligence trained by multiple reinforcement learning iterations. The technology exploits a neural network trained through reinforcement learning. At the beginning of each learning interaction, the quantum circuit is empty and is built iteratively through the interaction with the neural network, which decides which quantum gate append on the circuit. In particular, at each learning iteration, the network is given the information on the specific quantum algorithm to be compiled and on the low-level quantum logic gates that compose the circuit. Based on this information, the network decides which low-level logic gate to append on the circuit, choosing from a set of available low-level logic gates. At the end of the learning procedure, it is possible to recall the compilation strategy encoded in the network weights in a minimum time by compiling a generic quantum circuit.

Advantages

Training a neural network and encoding the solution of the problem in the network’s weights makes it possible to significantly reduce the execution times that could in principle be exploited for real-time quantum computation. Furthermore, by employing a reinforcement learning algorithm, the training procedure can employ an arbitrary set of low-level logic gates, making the compiler independent of hardware technology.

TRL

Team