search

desireevl / quantum-game-theory

Playable quantum game theory games
https://quantum-game.desireevl.com/
Apache License 2.0
67 stars 13 forks source link

Quantum GANs as Quantum Games #8

Open JansenZZ opened 5 years ago

JansenZZ commented 5 years ago

Training a GAN with one generator and one discriminator network is equivalent to finding the Nash equilibrium of a 2 player game. Can we extend GANs in the quantum setting to have competing generative networks and show that generative learning in a many-party setting is strictly more powerful in a quantum setting by some game-theoretic argument?

Some QGAN papers: https://arxiv.org/abs/1804.08641 https://arxiv.org/abs/1904.00043 https://arxiv.org/abs/1807.01235

desireevl commented 5 years ago

@JansenZZ @rajkk1 @elmo96 @GeorgeW95 all reading through these to decide if this is the path we want to go down for this project

rajkk1 commented 5 years ago

These two papers (one of which was already mentioned by Jansen) were put out on the same day and the authors mention each other in the acknowledgments https://arxiv.org/abs/1804.09139 https://arxiv.org/abs/1804.08641 Still reading through them though...

JansenZZ commented 5 years ago

More references:

Multi-generator GAN: http://rpubs.com/Firas/ganmg Multi-agent GAN: https://www.zpascal.net/cvpr2018/Ghosh_Multi-Agent_Diverse_Generative_CVPR_2018_paper.pdf

https://ieeexplore.ieee.org/document/6243378

This one has some rigorous results on quantum minmax games: https://arxiv.org/pdf/quant-ph/0207012.pdf

rajkk1 commented 5 years ago

So unfortunately I have many more questions than answers... I apologize in advance for my confusions!

  1. So from my (very cursory) understanding, having multiple generators kind of adds a "batch" aspect to the training, and hence it causes the GAN to converge faster?
  2. The circuit for a qGAN looks like this: image Does this translate to the EWL protocol (the one we've been using in our work so far, with the 'entangling' and 'unentangling' gates?) or would we have to change that up?
  3. We are trying to optimize the minimax game, and the original paper on qGANs already has a method to do this. Would our contribution be that we are doing this for a multi-agent qGAN?
  4. Are ~20 qubit quantum computers enough to implement a qGAN for the experimental part of the paper?

I really like the idea of combining quantum game theory with quantum GANs! I'm just trying to figure out, at least at a high level, how they fit into one another and what we should be looking for to contribute something new...