With very clear explanations in such a short time, why is this series not as popular as Sebastian Schuchmann's? Pretty jarring to see only hundreds of views on the same channel as a 15k viewer vid. Would you continue this series?
Unfortunately I don’t think so. I did make some progress on the game later, like adding some difficulty settings and menus… but nothing to make videos about. Like you pointed out, the views are kind of discouraging given the effort that goes into all this. I might pick up a different project in the near future and make a series about it though! Thanks for the continued support!
I find it useful to keep the last N agents when playing self-play in order to receive a more diverse experience. It's possible that agents learn a certain counter-strategy that only works against the current agent, but not against others. By playing against the last N agents, we should be able to find a more general strategy, rather than exploiting the weaknesses of a specific agent.
Yes! That’s what ML Agents’ ‘window’ parameter does… with a configurable number of steps, the opponent is changed by picking from one of the N recently saved agents! For this project, I kept it at 10 or something. Like you said, that helps with the agents learning to do well against a diverse set of strategies, assuming the candidate models were saved sufficient iterations apart…
Hello. I am currently working on a project for my diploma thesis using ml agents selfplay in unity. The game I'm trying to create is custom mode called "warlock" in Warcraft3 game. (but simplified - only shooting fireballs and pushing each other into lava, last player alive wins the game. (happens in 2D space (isometric))). I am however struggling and not sure why. I think my reward structure is fine as well as observations collected. I'd be very grateful for a consultation, if you could find the time (willing to pay a little even) to share your thoughts on it. I'd become "neural programmer", but it seems, there is not a specific video, that would help me so I'd rather do one time payment for a consultation (if you'd have me ofc). :)) really great video. (y)
This is amazing, thank you so much for the upload
With very clear explanations in such a short time, why is this series not as popular as Sebastian Schuchmann's? Pretty jarring to see only hundreds of views on the same channel as a 15k viewer vid. Would you continue this series?
Unfortunately I don’t think so. I did make some progress on the game later, like adding some difficulty settings and menus… but nothing to make videos about. Like you pointed out, the views are kind of discouraging given the effort that goes into all this. I might pick up a different project in the near future and make a series about it though! Thanks for the continued support!
I find it useful to keep the last N agents when playing self-play in order to receive a more diverse experience. It's possible that agents learn a certain counter-strategy that only works against the current agent, but not against others. By playing against the last N agents, we should be able to find a more general strategy, rather than exploiting the weaknesses of a specific agent.
Yes! That’s what ML Agents’ ‘window’ parameter does… with a configurable number of steps, the opponent is changed by picking from one of the N recently saved agents! For this project, I kept it at 10 or something. Like you said, that helps with the agents learning to do well against a diverse set of strategies, assuming the candidate models were saved sufficient iterations apart…
@@avb_fj... and the model hasn't stagnated/converged over those snapshots...
This is extremely fascinating. As someone partially involved with this stuff its cool to get further into it!
Thanks! Glad you enjoyed it!
Gold and underrated channel, hope you get the recognition for your good work.
Much appreciated!
Hello. I am currently working on a project for my diploma thesis using ml agents selfplay in unity. The game I'm trying to create is custom mode called "warlock" in Warcraft3 game. (but simplified - only shooting fireballs and pushing each other into lava, last player alive wins the game. (happens in 2D space (isometric))).
I am however struggling and not sure why. I think my reward structure is fine as well as observations collected. I'd be very grateful for a consultation, if you could find the time (willing to pay a little even) to share your thoughts on it.
I'd become "neural programmer", but it seems, there is not a specific video, that would help me so I'd rather do one time payment for a consultation (if you'd have me ofc). :))
really great video. (y)