Quick update and question about on_won_round()

[sarkaria]

Hello again @flemming-n-larsen ,

Quick update. The migration is going great. I have successfully integrate my neural network code into a Python Bot and initial results are very promising! :-) I'm currently running about 25k battles per experiment. Which is likely overkill :-). In each experiment I'm testing increasingly larger neural networks.

One minor issue, for which I do have a workaround. Is that on_won_round() is still not being called. I recall I might have mentioned this a while ago. Is it possible that the handler is missing from the base bot.py class?

Thank you

[flemming-n-larsen]

@sarkaria Good question! I will investigate that. 👀

[flemming-n-larsen]

@sarkaria It seems to be broken for all Bot APIs (Python, Java, and C#)

[flemming-n-larsen]

@sarkaria
This should be fixed with the 0.36.1 just released.

[flemming-n-larsen]

@sarkaria
FYI: I did find yet another bug causing the TPS to run a bit slower in the GUI, than it have to be. So This issue will be fixed with the next version coming soon.

Regarding the turn timeout, you can use the turn_timeout property returning the the turn timeout in microseconds (1 / 1,000,000 second (not milliseconds!)

And there is a companion property named time_left which gives you the amount of time left in the current turn. You can use this to get a good idea of how much work you can perform with the neural net. 😊
But be careful not to get too close to the turn timeout, as the bot will miss it opportunity to send new commands (it's intent) to the server, and thus will "skip" its turn.

You should be able to figure out how many operations you can do per turn based on the two properties. So if you measure you can do 500 "operations/cycles" per turn in average, I would stick to do just 450 - leaving some time for the web socket to send the intend to the server and "commit" your turn. 😉

And note that different systems will allow a different amount of operations/cycles. So you could have a "automated" built-in benchmark you use to determine how long time a single operation takes on a given typen to figure out, how many operations you are able to do per turn, e.g. in the start of the battle. The turn_timeout time can be changed for the game/battle as setting, so this is a good idea anyways.

[sarkaria]

Hello again @flemming-n-larsen

May I get your help regarding turn_timeout and time_left? (I suspect you may know about this already 😊)

Using v0.36.1, I captured the console output of a very simple TestBot I wrote. I placed the following line into the run() method and all the event handlers:
print(f"⏱️<method> Turn timeout / time left at turn {self.turn_number} = {self.turn_remaining}/{self.time_left}")

The values I see are very similar regardless of the value of TPS. The following file is an example at TPS=10
TestBotConsoleOutput_TPS_10.rtf

Thank you.

[flemming-n-larsen]

@sarkaria

This clearly looks as a bug, and the Python API is still young, so I guess it has a bug?

I will investigate a bit later. And if it is a bug, I will fix it before next release. It could also by some system timing issue (e.g. OS or library specific timing issue), which might be hard to overcome. Let's see.

[sarkaria]

No worries 😊

[flemming-n-larsen]

@sarkaria
I have fixed the two bugs #191 and #192 with the release of 0.37.0

When trying to fix the turn_left I discovered a similar bug in the C# versions. So it was a win-win. 😊

[flemming-n-larsen]

I have created this bug report #191

[flemming-n-larsen]

Created #192 for the time left issue