henrixapp / bomberman-rl-keras Goto Github PK

Resulting is a 5x5 matrix as input adding the transformed 💣, so we can train to avoid them

Plus reducing the number of displays of own position to one

Confirm whether left/right/top/bottom orientation of gym and real game are the same.

We should port Threads to Processors in order to gain more computational power.

Don't forget to rebase!

We observed that A3C is not performing well, when run with 2 threads compared with 16. We assume that is caused by the design of A3C which requires multiple actors to work on global model. Mitigated by 0e73661 for a3c_coinman

Currently number of crates is dense, should be switched to more sparse approach --> more different experience

Similar to what has been done here one could test a trained agent on an empty gaming area and see how it behaves. As there are neither coins nor crates one would expect that bomberman runs around randomly, perhaps sometimes dropping a bomb.

Explore whether it is feasible to use Cython to accelerate our gym_bomberman.

Bomberman needs to learn more extreme situation

Evaluate variants of DQN: double DQN and dueling DQN. By default the parameters in the DQNAgent constructor are like this enable_double_dqn=True, enable_dueling_network=False .

Reintegrate and merge all branches into master