Deep Reinforcement Learning for Quantum Error Correction
Typ
Examensarbete för masterexamen
Program
Complex adaptive systems (MPCAS), MSc
Publicerad
2021
Författare
Becker, Marvin
Chiu Falck, Karl-Rehan
Modellbyggare
Tidskriftstitel
ISSN
Volymtitel
Utgivare
Sammanfattning
Recent advancements in quantum computing have supported and reinforced their
promising computational abilities which for certain problems far exceeds that of
classical computers.
However, one of the main difficulties in realising larger quantum computers lies in
the instability of quantum bits often resulting in erroneous states. While there are
schemes in place to encode logical qubit states, such as repetition codes, the very
quantum nature of these qubits makes it impossible to directly observe their quantum
states without causing it to decohere. To circumvent this, parity measurements
of localized groups of qubits, so-called syndrome measurements, can be performed
as is the case in the surface code.
The encoding leads to improved stability and error resistance of the encoded logical
qubit. The challenge of decoding the surface code in the presence of potential errors
still remains: Multiple possible error configurations can lead to the same observed
syndrome state, thus making the decoding task non-trivial. This has been addressed
with different approaches such as the utilization of decoder agents trained via deep
learning.
In this study, we deploy techniques from the realm of reinforcement learning such
as Deep Q-learning, Proximal Policy Optimization, and Hindsight learning to train
agents capable of decoding errors on the surface code. To better reflect a realistic
setting, we also acknowledge the possibility of erroneous syndrome measurements as
part of the training. The entirety of these syndrome measurements constitutes the
observable state and introducing measurement errors breaks the Markov assumption
which is central to many reinforcement learning algorithms and leaves us with
a partially observable Markov decision process (POMDP). To deal with this, we
introduce the dimension of time to our problem formulation by keeping track of the
evolution of both qubits and syndromes. Our work includes a brand new state formulation
for the syndrome evolution and utilizes different strategies to train agents
on this newly-formulated problem.
Beskrivning
Ämne/nyckelord
Deep Learning , Reinforcement Learning , Quantum Error Correction , Surface Code , Quantum Bits