Quantum approaches to Generative Adversarial Networks and congestion games
Rigetti Computing and the Commonwealth Bank of Australia (CBA) have developed a new quantum technique that could help address challenges in training Generative Adversarial Networks (GANs) as machine learning solutions to financial applications. Based on solving a “congestion game,” where each player’s reward is diminished when sharing resources with other players, the team believes the technique could apply to additional sectors such as road transport network design and vehicle routing problems. The team verified their solution using a simulator of a six-qubit quantum computer.
Financial institutions are exploring how GANs might solve problems such as identifying fraud, predicting network infrastructure issues, and developing robust trading and risk management strategies. GANs support these use cases by training a discriminator neural network in a way that is robust to “fake” data created by a second neural network. The second network is known as the generator and it attempts to accurately model observed data. The generator finds additional applications in producing synthetic customer, market and transaction data, allowing information technology solutions to be developed without risking customer information.
Rigetti and CBA addressed a common challenge in GAN training, where the process can fail to produce the optimal discriminative and generative networks because each is stuck in a suboptimal solution that can’t be improved without the other changing simultaneously — a phenomenon known as a local Nash equilibrium.
Independently, a team at the NASA Quantum Artificial Intelligence Laboratory (QuAIL) identified calculating the Nash equilibrium as a key challenge in machine learning for which quantum computing may apply. The question the Rigetti-CBA team set out to answer was: Can we find the globally optimal Nash equilibrium and so provide the best performance for these potential applications? Answering this question is an NP-hard computational problem and intractable to solve on classical computing hardware.
The results demonstrate for the first time how a quantum computer can be used to estimate the optimal Nash equilibrium of a discrete network game reminiscent of traffic flow on a road network. Simulating six qubits, the team identified the most efficient combination of player strategies given six possible paths and two players. This solution, the optimal Nash equilibrium, is the state that for a GAN maximizes the discriminatory power and generative accuracy of the model.
Rigetti and CBA anticipate that this novel contribution to the field helps answer this question in the affirmative, and will lead to new quantum approaches to using GANs and game theory to solve machine-learning problems across sectors. This is the second application published in
recent months by the Rigetti and CBA team, following work in portfolio rebalancing.
Mark Hodson, Brendan Ruck, Hugh (Hui Chuan) Ong, Stefan Dulman, David Garvin. Finding the optimal Nash equilibrium in a discrete Rosenthal congestion game using the Quantum Alternating Operator Ansatz. https://arxiv.org/abs/2008.09505
About Commonwealth Bank of Australia
The Commonwealth Bank (ASX:CBA) is one of Australia’s leading providers of personal banking, business and institutional banking and share broking services. With 17.4 million customers and a history spanning more than a century, the Group’s purpose is to improve the financial wellbeing of its customers and communities. The Commonwealth Bank is Australia’s
leader in digital banking and maintains the largest branch network across the country. Headquartered in Sydney, Australia the Bank operates brands including Bankwest in Australia and ASB in New Zealand. For more information on Commonwealth Bank, visit www.commbank.com.au.
