I am not sure what to do next at the moment. I am current browsing Knight et al., 2000. This is a good book to read but there is not enough stuff for me to use immediately.
Last night I had a look for another paper by Cangelosi. The paper is "Evolution of communication and language using signals, symbols and words," 2001. Cangelosi and Harnad, 2002, mentioned that Cangelosi, 2001, discussed emergence of syntax and compositional languages.
The first place I looked at was IEEE. The keywords I used were, "evolution of communication." The search returned the following papers.
Do not forget that many of these papers are also available at other places such as Citeseer and CogPrints.
Evolution of communication in a community of robots
Perlovsky, L.I.; Fontanari, J.F
2005
Evolution of communication using symbol combination in populations of neural networks
Cangelosi, A
1999
Evolution of communication in a genetic based multi-agent system: use wise resources
Enee, G. Escazut, C.
2004
Evolving Compositionality in Evolutionary Language Games
Fontanari, J.F. Perlovsky, L.I.
2007
Evolution of communication in a community of simple-minded agents
Fernandez, J. Casals, A
2004
I then had a look at ACM. There were relevant papers with the keywords, "evolution of communication," nor where there a paper by Cangelosi but there was a paper by Nolfi.
Emergence of communication in competitive multi-agent systems: a Pareto multi-objective approach
Michelle McPartland, Stefano Nolfi, Hussein A. Abbass
2005
I also tried SpringerLink. A few papers related to my research by nothing had a higher level of relevance.
A few days back I also had a look at papers by Pfeifer and Bongard. I have quite a collection of those papers.
I want to conduct an experiment where I can handle the small issues I have with Cangelosi and Harnad, 2002.
In addition, I want the experiment to have a much larger memory so that each agent records and stores all input signals. I want to see if there is an alternative to a neural network or at least if I can augment the neural network so that things like conceptual revolutions as described by Dreyfus, 1992, are possible.
One of the other things I would like to handle is a concept described in Dreyfus, 1992. The concept is where the agent does not know what the agent wants and only realises what the agent wants after discovering it. You see the agents had hard coded reactions in Cangelosi and Harnad, 2002. These agents somehow learnt correct action through supervised learning. I object to this method.
I think the work Cangelosi, Harnad, Nolfi would benefit from ideas about embodied agents by Pfeifer and Bongard.
I am getting so frustrated. I have all these general ideas but it is so hard to make it practical!
Cangelosi, 2001 is quite hard to read. The main idea is to distinguish between signals, symbols and words. It does mention a little about communication where an utterance is a combination of symbols.
The paper does show how to set up an experiment with agents, neural networks and genetic algorithms. I would like to imitate the paper by introducing ideas from Dreyfus and Pfeiffer and Bongard.
Here is what I am thinking. These agents are only learning so much. They exist in a simple, world and they are enticed to satisfy a simple fitness formula. What would happen in a sophisticated world? What would happen if there were no simple fitness formula?
What would happen if fitness were simply how well you did in the environment?
How could researchers design these agents to learn and to keep on learning?
One of the things that would be required is moving out of the abstract world and into one, which has a greater degree of realism.
In addition, the agents need a different control system. The ones used by Cangelosi and colleagues are simple and they fit the task. How can researchers augment these control systems to scale up?
Differences
I have been spending the afternoon writing up a document describing the differences between the two types of artificial intelligence research when I notice that the document was getting a bit long. A colleague suggested that I write the document so that I could clearly express the differences between what I see are the two main approaches to research.
Currently the document is about 13 pages when it is supposed to be about one page!
There is just so much that I want to say!
It is annoying to now go back and write a single page. I guess that the longer document can serve as a set of notes for the one page discussion. One page, I can write one page!
No comments:
Post a Comment