Emergence of Communication/Student Reactions

From IPRE Wiki
Jump to: navigation, search

Student 1

This was a wonderful paper; though I have many things I comment on at the end of this response, I want to make it clear that I very much enjoyed this paper: it had a clear experimental model, and addressed the issue very well.

In this paper, they presented a simple world with a set of robots guided by nnets. The robots had the ability to produce and receive signals, of varying intensity, which they used to augment their individual actions, towards maximizing their "score" (fitness) as a team. The nnets were evolved over a number of generations to achieve this behavior.

The memory of the robot (that it had a sensor that specifically tells it what it said at the last time step) seems very ad-hoc; designing a system with actual memory, like a recurrent net, might be preferable? I know that we've complained when a system tries to model too many things, but this is something I feel is important enough to model.

How were the genotypes for the nnets shaped? I'm interested in how to encode a nnet such that it evolves well.

The deprived and no-signal conditions seem massively unfair; the robots have no way, beyond use of signals, to know where other robots are, except as "obstacles", which could easily be walls. As such, they would seem ti have a hard time correlating their team score to anything in their individual actions. However, the observation of how they "tend to optimize both their individual and social/communicative behavior" is cool; it addresses this somewhat.

How did the robots manage an oscillatory signal? What was its frequency? Did the intensity of the signal affect the distance it could carry?

Student 2