How could AGI build its own communication protocol?

Question

Let's say we have the basic scenario where two AGIs of about the same intelligence (but not same origins/code/model) have to communicate as efficiently as possible to achieve a common goal. Now we could have 2 starting points for that:

1. Either all they have is a common communication bus (e.g. sound, light, radio, etc.) and instruments (e.g. transceivers) to support it, and they have to figure out the rest.

2. Or they are some kind of advanced chatbots, but since the human language is lacking a lot to be used as a highly efficient protocol, they will have to communicate with what they have, to build a proper one.

Would it be possible to somehow induce them to communicate, and try to figure out what each other "say"? How could this be done?

And a more abstract question is how could this protocol "look" like?

Answer 1

This is a purely theoretical question, currently in the realm of philosophy and speculative fiction. Nevertheless, it is an interesting question, and may be instructive.

If we use the standard definition of Artificial General Intelligence as automata with human level intelligence, then they could certainly devise their own communications protocols, just as humans have.

1) These automata are AGI so they are creative and resilient, just like humans, and 'where there is a will there is [sometimes] a way." Absent robotic capacity they wouldn't be able to build anything physical and would have to rely on existing communications infrastructures. If they had access to 3D printing and versatile robots, they could probably build something new, but this would be infeasible for anything that requires extensive capital outlay, unless the automata first acquired a major communications infrastructure firm or two.

2) It's always interesting (and usually quite entertaining) to observe chatbots conversing with each other, but if they were AGIs, NLP is just one of many functions, and I doubt they would bother conversing with each other in human language, since all data is ultimately reduced to a string. If they were smart enough to be deemed AGI, they would certainly communicate in-species with the most efficient protocol available, and probably optimize it further, if possible, or create a unique protocol for purposes of exclusivity.

If they are truly AGIs, you wouldn't ostensibly have to induce them to communicate, because they'd be smart enough to understand the benefits of communication and cooperation, and would likely seek to form coalitions as a natural survival function. (Game Theory provides a mathematical basis for this.)

Multi-agents systems can self organize, even where the intelligence of the given agents is low, and in your scenario, the automata are smart.

Answer 2

1. I believe that some work by Randall Beer in the 1990s demonstrated that even simple agents could learn a shared communication protocol:

Computational and dynamical languages for autonomous agents

2. Viewed in general terms, a sentence uttered by an agent is just another 'feature' of the environment, from the perspective of an observing agent. Learning to extract and operate upon a subset of features is what modern machine learning is mostly concerned with.

While Braitenberg's book 'Vehicles' doesn't deal specifically with language acquisition, it does nicely illustrate some general principles for self-organisation of more complex recognisers and behaviours.