Reproduce result in paper

[franktseng0718]

I'd like to reproduce the result in your paper, but after I run the code this repo based on these settings(seed=1, reward_type=op, episodes=3000), the result in 1458 seems similar of the paper, but the result in 3427 seems inferior than the paper a lot. It seems even inferior than HB method. Below is my result.






The click value fluctuates a lot even in train in 0.25, 0.125 setting. I wonder is it normal? And how can I get the result in paper?