Journal 12/03/08

One of the things I would like to figure out is how to scale up research such as that done by Cangelosi, Harnad, Nolfi, Pfeifer, Bongard and others. These researchers have done great work on how language can emerge from simulated embodied agents.

Although the agents have been able to connect symbols to things themselves, researchers have hard coded basic behaviour researchers have been responsible for supervising the learned behaviour. I wonder if there is an alternative form for the control system.

For example, let say we expose an agent to a whole bunch of inputs including vision, touch, awareness of muscle contraction and hearing. The agent is also able to make noises as it fire out pseudo random impulses to test out its body. The agent also has to cope with the environment to take care of its body.

Now research has shown that simpler experiments with simpler agents have resulted in agents working out how to communicate together to make up their own simple language. In addition, research has shown that in another experiment agents were able to form categories and not only that, form categories by combining existing categories.

Now the question is how can this be scale up? How can we set up the situation so that the categories keep developing similar to the way humans have categorised their physical environment and events, which are non-physical. For example, the term chair can refer to a physical object and the term sit refers to something that is not an object but a property of an object. In other words, sit is a property of a person sitting.

That is one direction for extension; another extension is how can we set up the situation so that agents use the same word to refer to different things depending on the context? For example, chair may refer to an object or the action of mediating a meeting. This is admittedly an advanced concept and I do not expect that simulations of amoeba like organisms would be having meetings.

Moving on, another extension is multi word utterances. For example, an utterance comprised of the word sit and chair.

Yet another extension is how to train these organisms to transition over to using English like terms.

Back to the original query, how can we get an AI to coordinate all these inputs? I do intend to get to the stage where I can set up an AI on a single machine and let the AI record all inputs so that at appropriate times the AI can go back and re-evaluate the generalisations it has made but how can I get an AI to correlate anything? The AI could try different things to see what works and it could play around with the correlating algorithm as well. The AI could attach symbols to things, episodes and so on to reduce the amount of processing required.

Therefore, it is kind of a problem of learning to use many inputs to decide what is best in a situation when you do not really know what is best, and it is all undefined.

There is a description of some unsupervised learning techniques on Wikipedia.

Perhaps what I could do is clarify what I want the agent to do and see if the architecture that researchers have used is sufficient.

Yes! I have finally written my first draft of a roughly one-page description of what I see is the two main approaches to AI research. There are many similar discussions but writing this has helped me clarify what I am doing.

Yes! I have finally written my first draft of a roughly dozen-page description of what I see is the two main approaches to AI research. It is long, boring, repetitive, disjointed, biased and unreferenced but it is a completed first draft!

I have spent most of the day learning about unsupervised learning. I did not get very far today, it is a tad difficult.