I started exploring a new reinforcement learning approach for the 1024 project that previously reached a dead end using genetic deep learning on a simple convolutional neural network.
I'm testing a different training strategy (methods include DQN/PPO/A2C) to see if the project can be solved. I was able to reach a score of 1024, instead of the
theoretical limit of 65536.
If the approach proves successful, I'll publish a detailed write-up covering the design decisions, implementation, results, and lessons learned.
Episode Average Score Last Duration Max Duration Score
36900 754.98 807 1232 1024
37000 613.9 219 1232 1024
37100 619.86 380 1232 1024
37200 664.64 514 1232 1024
37300 667.97 498 1232 1024
37400 782.15 279 1342 1024
37500 688.46 220 1342 1024