Check out the next video of this series where I teach the agents to beat increasingly more difficult laser obstacle courses: th-cam.com/video/GV--pmWfIWQ/w-d-xo.html
As you found out towards the end of this video, a tuned reward function comes second to making sure the ai's are receiving all relevant information. scaling the rewards is a good thing to learn, and avoiding "absolute" observations like x-y coordinate positions are also good practice, unless you are trying to train the ai on a very specific obstacle course. One recommendation, if you do team sports again, is to replace the observations of "i am on the blue team, the one next to me is on the red team, it has a red ball" with "the one next to me is on the rival team, it has a ball of the rival color". That way instead of learning a strategy for both blue and red depending on the one input of which team they are on (so four strategies in total), they can just develop 2 strategies: one for allied objects and one for rival objects. Move the calculation of whether something is a team-mate or not to a simple binary check, instead of forcing the neural network to calculate it on top of everything else. it only needs to know its team color if you are doing something with asymmetric team abilities. Random idea for next video: archery. As opposed to dodgeball, where they have to find their weapons, they instead have a bow and arrow on them at all times, but each time they get a hit the target moves randomly on the x-y plane. they would have to control both their rotation towards the target but also the power they shoot the arrow with (i'm assuming the more power, the higher the arrow goes upward before falling back down.) Very good video overall!
Good suggestions! Speaking of archery, I was reminded of arrow tag, where the dropped arrows can be reused later on, giving a battle royal feel Speaking of which, dropping MLAgents in an actual battle royal environment sounds awesome
What if, instead of duplicating the arenas, you let all those agents fight in 1 big arena? Kinda like Zuzelo's stuff, but dodgeball. Imagine the chaos... (also ez content material reuse :) )
Huh so turns out Zuzelo's made a dodgeball video shortly after this, but sadly he also duplicated small arenas rather than his signature colossal warfare. Maybe it's a niche to fill out? :3
![SAFEPLANET - ดาวเคราะห์ [ THE PLANET ]](http://i.ytimg.com/vi/4X1XL3I_k3c/mqdefault.jpg)
Check out the next video of this series where I teach the agents to beat increasingly more difficult laser obstacle courses: th-cam.com/video/GV--pmWfIWQ/w-d-xo.html
BRO THE WHOLE TIME I THOUGHT I WAS WATCHING A POPULAR TH-camR'S CONTENT
Why hasn't this gotten more views and why do you still have less subscribers
New algorithm I guess. Earlier this year, this channel probably would've blown up like EightLittleBears
As you found out towards the end of this video, a tuned reward function comes second to making sure the ai's are receiving all relevant information. scaling the rewards is a good thing to learn, and avoiding "absolute" observations like x-y coordinate positions are also good practice, unless you are trying to train the ai on a very specific obstacle course.
One recommendation, if you do team sports again, is to replace the observations of "i am on the blue team, the one next to me is on the red team, it has a red ball" with "the one next to me is on the rival team, it has a ball of the rival color". That way instead of learning a strategy for both blue and red depending on the one input of which team they are on (so four strategies in total), they can just develop 2 strategies: one for allied objects and one for rival objects. Move the calculation of whether something is a team-mate or not to a simple binary check, instead of forcing the neural network to calculate it on top of everything else. it only needs to know its team color if you are doing something with asymmetric team abilities.
Random idea for next video: archery. As opposed to dodgeball, where they have to find their weapons, they instead have a bow and arrow on them at all times, but each time they get a hit the target moves randomly on the x-y plane. they would have to control both their rotation towards the target but also the power they shoot the arrow with (i'm assuming the more power, the higher the arrow goes upward before falling back down.)
Very good video overall!
Good suggestions! Speaking of archery, I was reminded of arrow tag, where the dropped arrows can be reused later on, giving a battle royal feel
Speaking of which, dropping MLAgents in an actual battle royal environment sounds awesome
I love these videos, keep it up mate. You rock
bro wtf how did this show up this is so good
Kindergarten gym class flashbacks
What if, instead of duplicating the arenas, you let all those agents fight in 1 big arena? Kinda like Zuzelo's stuff, but dodgeball. Imagine the chaos...
(also ez content material reuse :) )
Huh so turns out Zuzelo's made a dodgeball video shortly after this, but sadly he also duplicated small arenas rather than his signature colossal warfare. Maybe it's a niche to fill out? :3