Improve MCTS

[AlexanderNorup]

Improvements include:

• Fixed reward function to no longer return Infinity if the move is too good
• Updated custom-battle results summary screen
• Knowledge states! It's a setting!
• An updated function to help decide what to do. It is now based on turns till predicted wins with reward, instead of the best-child selection function used during treesearch
• You can now press CTRL + DELETE to quickly clear the Custom Battle screen
• You can now hold K in the custom-battle screen to randomize your monster
  • Hold CTRL during the randomization to randomize all monsters on both teams

Wow! That's a lot. And even without dirty hacks to make it work™.

[VictorABoye]

Cool stuff!
We could also look into trying GRAVE or other algorithms instead of UCT, if we want to try to improve it further.
loafs around

[AlexanderNorup]

Come to think of it, this choosing method has a fatal flaw. Consider the following scenario:

MCTS has two monsters and the opponent has two monsters.

Let's say that MCTS's active monster can win in 6 turns by doing continuous small amounts of damage, and MCTS's non-active monster has a move that will cause MCTS to win first turn it's used.. And let's say the opponent's active monster can't really deal much damage to MCTS's active monster, but would deal massive damage to MCTS's non-active monster.

In this case, this new method will be flawed.
MCTS will have the following options:

• Keep attacking with current monster (win in 6 turns, reward=10)
• Change to other monster (win in 2 turns, reward=0)

If you expand the "change into other monster", you'll see the following scenarios:

• The opponent uses the deadly move and MCTS loses (reward -100, win in turns=MAX_INT)
• The opponent switches to it's other monster (reward : +100, win in turns=1)

This would be a terrible choice, because of course the opponent isn't just going to give you a free win. It will try to kill you immediately.

Even through random playthroughs, it'll probably figure out that going down this path is pretty stupid as it looses in most of the time. but since this new system prioritizes "win_in_turns", it will choose this path anyway. So if you purely looked at rewards it would (in this case) do the right thing. But that dons't go for other cases like two gods on a single team.

Back to the drawing boards for new ideas I guess.

[VictorABoye]

In order to utilize the new battleresults from this in #18, I've decided to revert d63dc5a because of this issue.