The most famous conversation between two AIs was in the movie 'Colossus - The Forbin Project'. Colossus learns to communicate with his counterpart Guardian.

https://en.wikipedia.org/wiki/Colossus:_The_Forbin_Project

A real experiment was conducted by Luc Steels, Talking Heads. 'Robots' invented a new language. But it was rather an art project.

http://langsci-press.org/catalog/book/49

Language is a means of transferring a data structure from one mind to another. We know how to do that technically. In machine to machine communication this is called serialization / deserialization. Lisp expressions or XML are sufficient as a grammatical frame. Inflections aren't necessary. Data structures of an AI system can be objects in the sense of OOP. So the task of inventing a language can be reduced to agreeing upon objects, object property names and object property values that have their mapping for each participant.

Related:

https://www.fastcodesign.com/90132632/ai-is-inventing-its-own-perfect-languages-should-we-let-it

https://portal.ikw.uni-osnabrueck.de/~neurokybernetik/projects/alear.html

https://phys.org/news/2011-05-robots-language.html

http://www.itee.uq.edu.au/lingodroids

https://www.researchgate.net/publication/6378575_The_emergence_of_simple_language_in_an_experimental_coordination_game

http://www.economist.com/node/5135495

http://www.pnas.org/content/104/18/7361

https://www.eurekalert.org/pub_releases/2009-02/uoc-dcu020309.php

http://www.sciencedirect.com/science/article/pii/S0149763416301105?via%3Dihub

Regards,

Joachim

Dear Dibakar,

Thank you for your note

That much I know. These systems are, however, capable of learning. The more complicated they are, the better. Therefore I believe that AIs are (or will be) something more than a set of switches. This discussion, however, may lead to more philosophical questions, whether the may be a "ghost in the machine" in AIs? I know that some scientist consider humans too as complicated machines, but I sincerely doubt.

Dear Gyorgy,

A multi-agent system (multiple ' AI's' ) fall into one of two categories: centralized or decentralized control based systems. Centralized control systems consist of agents that have a certain degree of autonomy but the overall system is controlled by a unifying strategy to achieve a specific goal.

Decentralized systems differ from their centralized counterparts by having agents (AI's) with a higher degree of autonomy, but lack a pre-existing strategy that guides all of the agents. They typically have some form of communication system that allows the agents to develop the needed overall strategy while exploring their environment.

The communication or coordination between such agents can take many forms. One example of a Machine Learning system in which cooperative agents learn to exchange text messages (human language) is "Dynamic Scripting with Team Coordination in Air Combat Simulation"

https://reports.nlr.nl/xmlui/bitstream/handle/10921/927/TP-2013-507.pdf?sequence=1

The specific learning mechanism used in this example is called Dynamic Scripting, which is a form of (rule-based) Reinforcement Learning (learning through reward and/or punishment).

Hope this helps a bit.

Best regards,

Jan Joris

Dear Jan, thank you for your informative answer

The most famous conversation between two AIs was in the movie 'Colossus - The Forbin Project'. Colossus learns to communicate with his counterpart Guardian.

https://en.wikipedia.org/wiki/Colossus:_The_Forbin_Project

A real experiment was conducted by Luc Steels, Talking Heads. 'Robots' invented a new language. But it was rather an art project.

http://langsci-press.org/catalog/book/49

Language is a means of transferring a data structure from one mind to another. We know how to do that technically. In machine to machine communication this is called serialization / deserialization. Lisp expressions or XML are sufficient as a grammatical frame. Inflections aren't necessary. Data structures of an AI system can be objects in the sense of OOP. So the task of inventing a language can be reduced to agreeing upon objects, object property names and object property values that have their mapping for each participant.

Related:

https://www.fastcodesign.com/90132632/ai-is-inventing-its-own-perfect-languages-should-we-let-it

https://portal.ikw.uni-osnabrueck.de/~neurokybernetik/projects/alear.html

https://phys.org/news/2011-05-robots-language.html

http://www.itee.uq.edu.au/lingodroids

https://www.researchgate.net/publication/6378575_The_emergence_of_simple_language_in_an_experimental_coordination_game

http://www.economist.com/node/5135495

http://www.pnas.org/content/104/18/7361

https://www.eurekalert.org/pub_releases/2009-02/uoc-dcu020309.php

http://www.sciencedirect.com/science/article/pii/S0149763416301105?via%3Dihub

Regards,

Joachim

Thank you for the interesting comments. I try to understand :)