Unity ML Agents Having Constant 0 Cumulative Reward on TensorBoard

[deft marsh]

trying to learn ML agent toolkit by making a flappy bird ai, as you can see from the screenshots, the demo is recording rewards properly, the code is also properly executed (debug.log will execute when bird passes checkpoint), but when I try to train the brain with the demo i recorded, the reward is constant 0 as you can see from the tensor board screenshot

[bold ruin]

you are reseting your agent when they fail right?

[deft marsh]

yes

[deft marsh]

this is how i reset it

during demo recording it all works fine, but not during actual training

[bold ruin]

if the agent dies, you need to call EndEpisode()

that resets the agent itself and prepares it for another segment of training

call it whenever it hits a pipe

[deft marsh]

oh yeah i did that too

as u can see from the 2nd screen shot, the demo had 3 eposides and was getting all the stuff correctly

[bold ruin]

Well if your agent is improving and getting rewards, something could just be messed up in the results folder

Where tensorboard is getting the info from

[deft marsh]

is not learning at all tho, the eposide length is not increasing either

[bold ruin]

Send the scripts you are using

And a picture of all the agent scripts (your agent, behaviour params, decision requester)

[bold ruin]

Any reason the addReward() is commented out?

[deft marsh]

so orginially i was giving rewards not only on pasting through the pipe, aka point scored, but also was rewarding the agent for surviving, and i realized the bird was only getting the reward for surviing and not scoring points

so i commented out all the addReward() and just left the one for the point scored

[bold ruin]

should probably use addreward instead of SetReward, but I dont think that is your problem

have you tried print statements

like you are 100% that the collision is correct

and the code is running

[deft marsh]

i tried both, both not rewarding

[deft marsh]

yes

i have put Debug.Log() right before and after the addReward()

and both registered

[bold ruin]

okay, can i see your yaml file?

[bold ruin]

looks like you are using curriculum and/or demonstrations? personally, i dont think you would need to approach this in that kind of way, so it could be something to do with that

i don't have any experience with curriculum or demonstrations so i wont be able to help you in that case. your yaml and code both look fine so

[deft marsh]

im not using curriculum, is just the name

one of my project memebr was naming things like that so i didnt bother to change

if u are referring to line 2 that is

i am using demo tho

[bold ruin]

yes i was. I am also talking about the demonstration recorder attached to your agent object. i have a feeling that it is messing up the rewards or you arent implimenting rewards properly. Have you read the documentation or only the tutorial?