Some problems about collecting data and model

[immortalsplay]

Hey, Tea Pea Thank you for your excellent work! Me and my team inspired by your work and try to use it in Retake Mode. There are some questions:

1. We try to collect data from some experts. So the “hacking”(read memory) way is not available. But when we used Win32api collect data we found CSGO always renew the mouse place. Can’t get stable data. Did you have that question before?
2. We try to use decision transformer in this model but the ai doesn’t work well (states & actions input). Did you used RL or transformer when you training? Do you have some suggestions to improve the performers? Looking for your response.

[TeaPearce]

1. Yes I had big trouble accurately capturing and sending mouse events. The RAM reading was the only was I found to work unfortunately.
2. I did supervised BC, but using CNNs (+ lstm). I suppose the main idea should be similar if switching to transformers, but implementation might bring some issues.