Another season and my bot finally made it into Champion I. 150 reward cards and ELO of 5277.
Improvement explained
At the start of last season I switched to a regularization approach for my Monte Carlo Tree search. Before that my approach was as follows:
- Pick a team through tree traversal
- Evaluate this team against n teams and sum up the reward function values
This approach had two issues: with larger n the MCTS could do less iterations because it was using more time on simulations. Plus rewards could cancel each other out and overfitting might take place. Bad design in general.
My new approach:
- Pick a team through tree traversal
- Evaluate this team against one random team from a pool of n teams
Now the MCTS isn't stuck as much in simulations and can do more iterations. It should also counter overfitting on one specific team in general.
With that change aicu managed to play in the lower Champion I league. Somewhere between rank 90 and 70. Aside from that, I bought the remaining legendary summoners at max level (including Archmage Arius) at the end of this season, which gave aicu a little push as well.
Sure, Aicu is playing now with a 95% max level deck, but how did the changes affect aicu-chan a deck with a couple reward cards and only level 1 Summoners? Well, it pushed all the way to Silver II from Bronze I last season. The corresponding chart is in aicu chan's update post.
Improvements?
Further improvements are starting to get more complicated. For now I'm trying to figure out if I want to follow the "intuition" approach of AlphaZero or try some other things along the lines of temporal difference learning. Need to brainstorm that a bit, the approach of AlphaZero shows great promises in Hearthstone, Go, Chess etc. But I'd actually like to experiment with other approaches before I go down that route. The Trie approach which I explained last time is something I want to explore further.
Giveaway
 |
|
 |
|
|
| Naga Fire Wizard | Naga Fire Wizard | Javelin Thrower | Javelin Thrower | Javelin Thrower |
For this giveaway there's less variety but more Earth and Fire Splinter awesomeness :). Aside from that the same rules apply as last time. Everyone who comments below with a valid SM username get's added to the giveaway pool. Cards are distributed at random at the end.