search

berickcook / AIRIS_Public

AIRIS Public Release
http://airis-ai.com
GNU Affero General Public License v3.0
12 stars 3 forks source link

Learning casualty of independent actions / switch to ONA #3

Open kripper opened 1 month ago

kripper commented 1 month ago

Hi.

Very nice video. Congratulations.

I wonder why you didn't implement observing and learning casualty of actions that are independent from the AI player bot (the actions of the space invaders). You mention this at the end of your video.

In general, it will be too resource intensive to feed all this environment variables into the NARS engine.

What does NARS offer to implement an attention mechanism to filter those environment variables that are worth exploring?

And how does NARS avoid exploring over and over the same environment variables that have proven to be useless (because no casualty relation could be inferred with enough casualty).

kripper commented 1 month ago

Oh, wait! You are not using ONA. Why not? Have you considered switching to ONA.

berickcook commented 1 month ago

Thank you! I've been trying to get it to learn independent causality, but it is a very challenging problem. Separating causal relationships is very tricky.

As for ONA, AIRIS is actually more capable. I work closely with Patrick Hammer (One of the lead developers of ONA), and he is working on his own implementation of AIRIS called NACE. He is incorporating relevant parts of NARS into AIRIS to see if we can overcome the independent causality problem, but so far we're both still struggling...

kripper commented 1 month ago

Can you please elaborate more in detail about why is it different to observe and identify causality on other actors?

Does NAL7 help here?

berickcook commented 1 month ago

Sure! When the agent performs an action it is associating the changes that happen in the environment to that action. When other changes are occurring at the same time due to other actors those changes get mixed in with the actions of the agent. So the agent thinks its actions are the cause of the changes caused by the other actors.

Ideally, with more observations, it should be able to dissociate those changes. However, we (Patrick and I) haven't found a stable mechanism for that yet. We have had some success, but there are several edge cases that still trip things up. Such as a moving block that goes back and forth between walls.

As for NAL, Patrick would be better equipped to answer that as he is much more familiar with it than I am.

kripper commented 1 month ago

So the agent thinks its actions are the cause of the changes caused by the other actors.

I think that's fine initially, and that the confidence level should decrease with more observations.

Maybe the problem is that these observations continuously generate new wrong confidence level (noise) over and over again?

If so, it is important that the system learns to avoid rehashing the same failed ideas.

kripper commented 1 month ago

AIRIS is actually more capable

Can you please explain more in detail?