Artificial Intelligence Depot
Visiting guest. Why not sign in?
News, knowledge and discussion for the AI enthusiast.
FEATURES COMMUNITY KNOWLEDGE SEARCH  
Genetic Programming : An Introduction
Imagine a world in which computers program other computers based on strategies borrowed from biology and natural selection. The book explores such fascinating possibilities thanks to AI!
More information at Amazon US UK

Reply to Message

Not registered yet?

The AI Depot has a focused community of friendly users. Rather than let anyone abuse the site at the brink of promiscuity, we prefer to let only those with an active interest participate... this simply requires registering.

Why not sign up!

Joining the site's community is completely free. You can then post messages freely, and customise your personal profile at will. Specific privileges will also be granted to you, like being able to access printer-friendly articles without restrictions. So, why not register?

Username:
Password:
Subject:
Email me when someone replies.
Body:

Parent Message

Behaviour-Based AI

God, you've got to stop posting stuff like that! You're making me really want to have my own robot ;)

Emotions are a great way to create very interesting emergent behaviours using very simple techniques. In your case, you could model a few emotions and let the bot learn what makes him feel more comfortable... For example, when there's lots going on and he can't understand, he'd feel scared and confused, thereby eventually learning by negative feedback to switch off.

The behaviour mechanism you described is very popular among robotics researchers... surprisingly, it's called behaviour based robotics, whereby the robot needs to decide how to behave based on his current state. MIT Media Lab is doing a lot of research into combining all this with emotions (using artificial sheep-dogs for example). Very interesting stuff indeed.

935 posts.
Saturday 08 December, 09:33
Reply
Machine learning

The MIT Kismet robot is probably the best example of vision-based social interaction and modelling of emotional states. This is obviously a lot more sophisticated than Rodney, but the basic principle is the same.

To demonstrate a learning algorithm it would be possible to have Rodney's motion tracking ability "emerge" using a reinforcement learning system combined with modelling of emotional states (sometimes refered to as "value systems"). However, for this type of behavior conventional logic suffices quite adequately and the use of RL would probably be overkill and degrade the robots realtime performance. In my opinion its better to reserve these kinds of adaptive system for more difficult tasks such as identifying and classifying *what* the robot is actually looking at. For this task I've used a Kohonen-style SOM.

The SOM actually performs very well indeed and if anything learns too rapidly, such that I've had to turn the learning rate and neighbourhood size right down. If the learning rate is too high the SOM is in a continuous state of flux and never really settles into good classifications for any reasonable duration. After some period of time I could turn off the learning and manually assign classifications to areas of the topological map, such that the robot could then make logical inferences about objects being observed and the sequences of observed events.

- Bob

136 posts.
Saturday 08 December, 10:22
Reply
Reinforcement Learning and SOFM

I'm actually just looking into RL for my bots. There's quite a bit to be done, especially for speeding up the learning. But as you said, it does have an impact on the performance, and you also notice some fairly erratic behaviour at first! This can be quite dangerous for a physical robot, but taking especial care when modelling the problem seems to do the trick...

One thing you might be interested in for learning feature maps, are Neural Gas Networks. They're quite popular at the moment, and rightly so! They can expand to match the dimensionality of your data, and refine and contract as more feedback is obtained from the environment. I'll see if I can find some links for you...

935 posts.
Monday 10 December, 08:36
Reply
Hyper

For a physical robot its definately useful to have the outputs from whatever adaptive system is being used put through a checking routine to ensure that the desired servo positions won't cause the robot to be damaged.

The ability to have the feature space expand or contract depending on the dimensionality of the sensory input received by the robot would certainly be an advantage. One problem with using a SOM type system is that according to the classical algorithm the system gradually moves into an equilibrium state (learning rate and neighbourhood size gradually reducing at some fixed rate) until a solution of some sort is reached. Of course in the real world living systems never reach a state of equilibrium, but are instead in a continuous state of change. You could argue that the ultimate equilibrium state for a living system is death, but even in death biological systems are in a state of being degraded and eventually transformed into other types of organic matter or new forms of life.

I recently upgraded my PC from windoze ME to XP, and there does seem to be a dramatic improvement in the performance of the video cameras, to the point where the gains which I was using for motion tracking no longer work. With much faster video processing the robot appears to become hyperactive, with the servos perpetually overshooting and trying to compensate back. It's probably worth making the gains adjustable in the software so that it could potentially be run on different PCs.

- Bob

136 posts.
Monday 10 December, 14:24
Reply
Upgrades

I'm surprised and impressed that the OS upgrade made such a difference!

I guess adding parameters to adjust the gain was a good idea anyway, but I'm sure there are plenty other things that you could do with the extra speed...

935 posts.
Wednesday 12 December, 05:49
Reply

Back to the Artificial Intelligence Depot.