And with some delay my report on the last season. My main deck run on Aicu performed as usual and finished at 5288. Aicu-chan performed a lot worse.
Performance Explained
I experimented a lot with the MCTS algorithm this season and tried to add "more intelligence" by encoding game logic within the team selection. Didn't work as intended. So I had some hiccups in the performance and Aicu tumbled through the ranks. After reversing that, I let it run until season end and it reached a ranking of 5288.
The reason why my level one account aicu-chan performed so much worse, is because it got stuck in match searches. I think this was mainly caused by very little players in the lower ranks. I haven't caught this case, and I had to change the code for that. Which took some time, because I was busy at work. But now Aicu-chan runs stable again, and I hope it will reach silver league or higher again.
Future Work
Like I mentioned in the earlier posts, I think the basic Monte Carlo Tree Search (MCTS) has reached it's limits. So I either need to implement a proper neural network and use it to guide the MCTS. I've experimented a bit with bidirectional recurrent networks a bit on past match data, and it showed promises. But I honestly don't want to go down that route. Mostly because I think it's boring. Every major AI paper within the last couple years is: neural network this, neural network that. I do see that they perform very well when data is abundant, that's why I'll try to improve my curator bot with better language models like BERT and GPT-2. But with my Steem Monsters bot I won't be going down the Neural Network path yet.
I've come across a couple of interesting papers from last year describing Online Temporal Difference Learning algorithms as well as a generalized MCTS algorithm, which works on any Markov Decision Process. So essentially Monte Carlo Graph Search. That's what I'm going to implement next. Afterwards, I'll experiment with various neural networks to boost those approaches. If they need boosting.
Gold Legendary Anyone?
With the dwindling supply of Beta Packs, I bought 50 for one last time, along with Brilliant Gold and Legendary potions. The result is rather shiny:
A few legendary summoners along with a gold foil legendary summoner. (and many more gold cards)
Giveaway
This time the giveaway will be more compact. Just two neat little gold cards and something epic.
 |
 |
Comment down below with a valid Steem Monsters username and you will enter the distribution pool!