The fruits and bombs are seemingly predetermined and the same every time, so could you just use reinforcement learning until your AI could get to a score of like 100,000? Also could scanning the screen for non black or non brown pixels work?
My all time best was 11932 with the best ammount of pomegranate hits being 89, seeing an AI getting 600 with not perfect "technique" makes me cry imagining that in 1 year of development it could become so good that reaching 6 digits would be easy T_T
You can easily increase your model speed if you use lower screenshot resolution (not a game resolution), then turn image black and white, then use color contrast to make objects looks clear. And also, check to click on fruit little bit away from bomb.
If you wanted to improve the ai, you only need 3 frames and some calculus to predict the exact trajectory of the fruit meaning you'd be able to slice every fruit on the 4th frame that it's on screen
there is a significantly faster way of doing this which is looking for clusters of pixels that have changed and only using the AI to detect the bombs to make sure those are in blocked off zones for the motion detection
