1

Let your agent play some games - perhaps about fifty or so - on its own before you start teaching it. This will allow it to gather a basic vocabulary and give you an idea of the words in use at the moment. People from all over the world are taking part in the experiment, and the agents create their own words when necessary, so the language is constantly changing.

2

Different locations have different light conditions and different images. Words that work well in one place might not be so useful in another.

3

Remember that agents don't talk about specific objects, but about properties of the objects in the scene. The words that agents use are really adjectives, not nouns.

4

The agents play a guessing game. The words they use to describe the topic (the area framed in black in the image) must distinguish the topic from all the other objects in the scene. In the scene below, for example, the agent probably wouldn't use 'red' (or a word meaning 'red') to pick out the marked image, because there are other red objects in the scene. Words referring to the position, size or shape of the object might work better.

5

Words can be ambiguous (mean more than one thing) and there may be synonyms (different words that have the same meaning). Sometimes one group of agents might use a word to mean one thing while another group uses the same word to mean something clearly different. This can happen because one agent gets the wrong idea about a word when it first hears it. If it heard the word 'green' while looking at a scene in which the topic was one big green object surrounded by several small red objects, it might assume that 'green' actually meant 'big', and then communicate that to others. The result could be one group of agents for whom 'green' means 'green' and another for whom 'green' means 'big'. This kind of confusion is inevitable, but normally it doesn't last long; the self-organizing dynamical processes that govern the robot population damp out the confusion.

If you get bored of waiting for evolution to do its work, why not do something to speed the process up? Take one of your agents that knows the 'right' meaning of the word, and send it to a site where there are other agents that are using the 'wrong' meaning so that it can teach them. But wait ... if there are enough other agents there, maybe they'll teach your agent their meaning, and it will come back with the 'wrong' meaning! Maybe you'd better send two agents ...

6

The robots speak their own language and it changes over time, so words may not mean what you think they do; the meaning of the word 'blue' isn't necessarily the color blue. To the robots, the word 'blue' might mean something different from what we'd understand by the same word - it might represent some specific shade of blue, or perhaps something entirely different like a quality of shape or size. The only way to work out the real meaning of words in the robot world is by looking at the way that they're used.