Couldn't reproduce task2 of Gym-MiniGrid

[shinya16864]

Hello, thank you for fantastic RIM implementation.

I tried to reproduce the result of Task2 of Gym-MiniGrid in README.md, but I couldn't.

I made 10 different models (different models correspond to different seeds) in learning, and calculated their performance by following the instructions described in README.

I calculated episode rewards of 100 roll outs for each model (1000 episodes in total), and got mean and std of them.
As a result, I couldn't reproduce that RIM (Kt = 4, Ka = 3) is better than LSTM in MiniGrid-MultiRoom-N4-S5-V0 (4 rooms).
Mean episode reward of RIM (Kt = 4, Ka = 3) is 0.07, while that of LSTM is 0.27 .

Could you let me know your settings when you got the result of this task (seeds and others, if any)?