The goal is to explore the behavior of generative adversarial networks (GANs) in a setup with >2 agents. Initial exploration:
- one generator,
- playing against 2 or more discriminators,
- discriminators decide under consensus protocol
- consensus model on of: {majority, unanimity, min, max}
- Specifying discriminator population
- Coordinating decisions with discriminator population
- incl. coordinated training of discriminator population
- Generator: random selection of discriminator output for generator update?
- discriminator throughput combine (avg)
- select winning D(x)? (max)
- select losing D(x)? (min)
- noise injection
- need new math to justify...
- discriminator scoping id procedure is sub-optimal. non-unique names for panel sizes > 9
- GAN training hacks