Simplistic while functional 2D multi-agent RL
Really all I did here, is modify the simple World I created a few months back to support multiple agents (blue-dots).
Then I simply loop through each agent in turn (which means, they act sequentially), which is a simplification indeed, but sufficient for a quick demo.
Finally, I train them, 16K steps for good measure, to look for green dots (food) and avoid the rest (red enemies, walls, etc.).
Here is what it looks like (I even added an audio track :D)
What for?
Good question. Because I can? Nah…
But also, consider this: Each of these 5 agents here have learnt independently their own set of rules with the LCS algorithm. And they seem to do a pretty good job.
In terms of Applications, I don’t know, but maybe a non-centrally coordinated, fully distributed behaviour, for a cleaning team of robots would be a kind of application here.
And no, that’s not a great idea (it would be easy to teach the robots what to do in this particular scenario) here. But in a more complex setting, maybe writing down behaviour rules would be harder, and then the LCS approach might make sense.
Conclusions
I don’t kid myself, but I was half-hoping to see some kind of “emergent” behaviour :D Obviously, that didn’t happen.
Then again, there is much more to that, and my agents are pretty simple… Oh well.
It works. That counts.
Next up: I think I want to look into “Complex Adaptive Systems”. I have found a book that explains them pretty good, and provides some examples in prose, that I might want to transform into actual computer models… If, you know, I make the time for it.