An idea popped into my head (one of few, and probably not original) that I need to get down and have people chew over. A lot of people write talker bots that are intended to respond in a human manner, what about an eviroment for those bots to interact with each other, a sort of bot client. Instead of the bots having to wait for humans to teach them, they could be talking to each other and...I don''t know what would happen. But people could hop into the client at anytime, and interact with the bots. It would help to hide the identity of the bots, so you don''t know they are bots. Also, maybe we could set up a reinforcement system. If a bot responds too poorly, it is killed, and the other bots see this, and try to learn from it.

That sounds like a good idea and I wonder why it hasn''t been done in the past.

---
Brent Gunning | My Site

One of my friends set up an MSN bot which talked to other people who talked to it. He also set it up so that it could send the first statement. I forget which engine he used sadly though. But it was quite funny when we got them to talk to each other. They just talked about such random stuff!

- Rob Loach
OverTech Technologies
-----------
"Life moves pretty fast. If you don''t stop and look around once in awhile, you could miss it."
- Ferris Bueller

Yeah i''ve been thinking ALOT about that concept! I want to use evolution etc...the only problem is i kinda have to finish my current game first because im doing it for a class. NLP and AI in general is a really exciting area that hasnt been developed the way graphics etc have. I think eventually someone will invent an AI that will blow everything else away. Like the advent of the microprocessor.
About the specific idea...I think it would be neat to give each bot a personality. That way they would handle concepts differently and affect eachothers thought process. If you had two dry bots talking to eachother it would be like
"Hi"
"Hi"
"How are you?"
"I dont know what you mean."
"I dont know what you mean."
See what i mean? These bots must become much more dynamic to actually work...just some of my recent thoughts...not that its anything really new.

Brian

To a certain extent it has been done and the results were very impressive. I cannot recall off the top of my head the name of the principal researcher, but there was a world wide collaboration to develop bots that could talk to each other or a human. Each agent was a piece of mobile code that could jump to a different location. At each location there were two computers, each with a video camera aimed at a desktop with various items on it. At any particular time there were two agents at a given location, conversing about the environment and interacting with the human. The bots would spend a short period of time at a location and then jump to another location to converse with a different bot. If the bot didn't know a word for something, it created one from a list of phonemes (sp?). The other bot would then (possibly) learn this word from the context it was used in. It was really fascinating stuff!

The project even had a public stage where the general public could interact with the bots, but that was shut down because people kept teaching the bots expletives.

Apparently not only were the bots able to develop a language with several different dialects (that evolved in time and space) but the language they learned also contained a simple grammer, which was totally unexpected... there are some very interesting and controversial linguistic results that stem from this research and the fact that grammer can be learned!

There was an article earlier this year in New Scientist about the research, but I cannot recall which edition. You could try searching the online archive for the article or the web for more information on conversation bots.

Cheers,

Timkin

[edited by - Timkin on December 3, 2002 1:29:21 AM]

whoa...

Having done this stuff before, I can give a few remarks.

quote: . Instead of the bots having to wait for humans to teach them, they could be talking to each other and...I don''t know what would happen.

What happens is that the bots get stupid.

Ya see, a conversation bot needs, above all else, BREADTH of experience. It needs to know about football and cooking and what color white grapes are and how many eyes a fox has and how edible shoelaces are. What it does NOT need is to talk to other bots with small experience bases.

You sound perilously close to inventing genetic algorithms. (what you''ve got so far I''d call "genetic inbreeding algorithms" ). You might want to read some papers on them. While chatterbots are possibly the worst application for genetic algorithms ever, genetic algorithms might well be very useful for other problems. Spend some time on Google.

Don''t listen to me. I''ve had too much coffee.

How would you evaluate fitness without human intervention? If there was an algorithm which could judge good bot-speak from bad bot-speak, wouldn''t we just use that algorithm to generate the good from the start?

nah... just adding this algorithm would not help. there still needs to be added content, which bots without presetting it can not.

This has been done, years ago, except that most conversational systems do not change their fundamental performance with conversation, only their memory of past conversation (possibly associated with specific individuals). So, Racter might pick up new "material" from Eliza, but the way Racter works doesn''t change.

quote: Original post by capn_midnight
An idea popped into my head (one of few, and probably not original) that I need to get down and have people chew over.
A lot of people write talker bots that are intended to respond in a human manner, what about an eviroment for those bots to interact with each other, a sort of bot client. Instead of the bots having to wait for humans to teach them, they could be talking to each other and...I don''t know what would happen.
But people could hop into the client at anytime, and interact with the bots. It would help to hide the identity of the bots, so you don''t know they are bots.
Also, maybe we could set up a reinforcement system. If a bot responds too poorly, it is killed, and the other bots see this, and try to learn from it.