Congrats!

[CallMeBW]

Congrats to the first place in the Chicago Event! I have tried out your SelfMotivator code before and have just seen that you won Chicago! Well done. I guess you were able to create an even better model.

[scottpletcher]

That winning model was the SelfMotivator I tried in Atlanta but without the same outcome... just on that track, in that car, under those conditions, it worked well. Curious if others with physical cars can replicate the performance. It was really smooth and hugged the outside line.

[JenkinsJB]

Can you share any info about the hyper parameters you used in Chicago or Atlanta? How about training time?

[scottpletcher]

Sure. For the winning model in Chicago (codename "SelfMotivator") I trained it in SageMaker straight as that model was trained before the DR Console was available. I used the default settings from the repo that was published by AWS for DR training.

The default values in that version matches the default values in the DR console now. As for steering and speed, that repo had those items hardcoded in the backing python code, but they were a 30/15/0 angle breakdown and a 3.33m/s and 1.67m/s speed breakdown.

[loveunk]

@scottpletcher thanks for sharing! Will you attend the Hong Kong Summit? Hope to see you there.

[scottpletcher]

Thanks Kevin! I won't be at the Hong Kong Summit but will be at the New York City Summit in July.

On Sun, Jun 16, 2019 at 9:22 AM Kevin notifications@github.com wrote:

@scottpletcher https://github.com/scottpletcher thanks for sharing! Will you attend the Hong Kong Summit? Hope to see you there.

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/scottpletcher/deepracer/issues/1?email_source=notifications&email_token=AEVGPP7VKG4LNHOK2DSUUPTP2Y47XA5CNFSM4HR2T4HKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGODXZM5TQ#issuecomment-502451918, or mute the thread https://github.com/notifications/unsubscribe-auth/AEVGPP7CWFFJ65VYAGROOF3P2Y47XANCNFSM4HR2T4HA .

--

SCOTT PLETCHER INSTRUCTOR

Indianapolis, USA - +1 971 227 4667

www.acloud.guru

[Carsair]

@scottpletcher Great presentation, great repo. Thanks so much. It's been really cool to look while dusting off some of my notes from the Atlanta conference. Had a few questions if you had time. Did you run the same model in NYC?

It seems like using progress as the only motivating factor and allowing the reinforcement learning to do its thing could be even better than including the speed as factor. (sometimes slower can be better) Have you tried that?

Have you tried any of the virtual circuits?

Thanks!

[scottpletcher]

@Carsair in NYC I ran 5 different experimental models--two of which showed some promise. I really was looking at NYC as prep run for Vegas. I'll post them to github soon. They were variations of the Atlanta (Self-Motivator) model. I haven't spent any time at all on the virtual circuits mostly due to time constraints. That is an interesting idea about using just progress....we should definitely try that one out. However, the end objective is to have the fastest time and without some other time component, I don't think the model will learn the right behavior. It might be happy to roll along at 0.01 m/s, but racking up the progress reward. One thing that I've learned is the training process will definitely expose holes in your reward function through unintended behavior.

[pnhatminh]

@scottpletcher Great presentation, great repo. Thanks so much. It's been really cool to look while dusting off some of my notes from the Atlanta conference. Had a few questions if you had time. Did you run the same model in NYC?

It seems like using progress as the only motivating factor and allowing the reinforcement learning to do its thing could be even better than including the speed as factor. (sometimes slower can be better) Have you tried that?

Have you tried any of the virtual circuits?

Thanks!

@Carsair could you please be so kind to explain to me why using only progress parameter result in a better performance for the training?

[scottpletcher]

@pnhatminh Sorry about the delay in response...I guess I missed the notification. To answer your questions, in NYC, I ran 5 different models during my 4 minutes...just trying to figure out which experiment had the most promise. I don't have a car or a track so summits are the only place I get to run in a real car.

I've tried all sorts of combinations and one thing I can say is that results in the simulator usually differ with results in a real car. I've tried some of the virtual circuits but haven't given it much attention. It kind of seems there are two disciplines emerging...training for physical racing and training for virtual league just because of accessibility, performance and training process. I'll usually train lightly in the sim just to make sure I'm on the right path, then shelf the model until I can test it in a real car.