Reinforcement Learning for Multiple Agents - Superiority Discussion

[kartoun]

Hello!

 

I'm attempting to show a superiority of a collaborative agent over

other agents exist in a learning system. In such a system there is a

one unique agent that is able to learn both from interaction with the

environment and from the other learning agents that exist in the

system. The other agents can only learn from interaction with the

environment, independently and are not aware of the collaborative

agent or any other agents.

 

I don't know if it is possible to prove superiority without attaching

conditions to my proof. Thereby, I added a few conditions. Here is my

progress (it's short):

 

http://www.ie.bgu.ac.il/kartoun/site/ReinforcementLearningMultipleAgents.pdf

 

If I will be able to show that Equation 7 holds then the collaborative

agent superiority will be demonstrated.

 

However, to show that Equation 7 holds, first I have to show that the

residual in both sides of the equation are positive. If I show that (I

don't know how - maybe add another assumption?), then the next stage

will be to show that Equation 7 really holds (currently I don't know

how to show that as well).

 

Here is a convergence proof for a single Q learner as suggested by

[Jaakkola et al., 1994]:

http://www.ie.bgu.ac.il/kartoun/site/ConvergenceProofforSingleQ.pdf

 

I hope that someone can help/give an opinion/share some thoughts about

this issue, please.

 

Thanks and happy new year!

 

[fabrizio]

I read your email just now... It seems really interesting! An agent thar learns from others? woow awesome idea!

 

I trying to navigate through your link but after some seconds I receive an "internal error" ..

I'm a MSc student and I recently attended a machine learning course, I'm interesting in q-learning, here a blog with my project http://machinerysoul.blogspot.com/

 

I would like to read more about collaborative agents, could you give me some hints about it?

 

cheers

 

bye!