HearthSim — Tuning the AI

By | June 13, 2014

The HearthSim AI is controlled by various model parameters, and it is difficult if not impossible to find the optimal set of parameters that will perform well under all circumstances. So, we need to be able to break down the parameters and understand them in more manageable chunks.

For those of you unfamiliar with the AI model, see this post.

Today, let’s look at the board score weighting. In the model notation, they are \(w_{\rm a}\), \(w_{\rm h}\), \(\tilde{w}_{\rm a}\), and \(\tilde{w}_{\rm h}\).

Roughly speaking, these weights represent how much you value your minions and your opponent’s minions. If your own board score weights (the \(w\)’s without the tilde) are high compared to your opponent’s board score weights (the \(\tilde{w}\)’s), then you (the AI) is not likely to make trades with your minions; the value (the score) that you lose by losing your minions would be much higher than the score you gain by killing your opponent’s minions. On the other hand, if your board score weights are low, you (again, the AI) are likely to make aggressive trades, as such moves will help to maximize your score. In other words, the difference between the opponent board score weight and the self-board score weight determines how aggressive / control-oriented the AI is.

Here is a numerical experiment. We take two players playing standard minion only decks. Player 0 is going to be controlled by an AI with different weights \(w_{\rm a}\) and \(w_{\rm h}\), while player 1 is going to be controlled by the standard AI with \( w_{\rm a} = w_{\rm h} = \tilde{w}_{\rm a} = \tilde{w}_{\rm h} = 1\). Under this circumstance we record the win rate of player 0. For the save of brevity, we will take \(w_{\rm a} = w_{\rm h}\) and call them “aggressiveness.” Note that the higher the aggressiveness, the less willing the AI is to trade minions with the opponent instead of going for the face.

The results are summarized below:

Aggressiveness P0 Win P0 Loss P0 Win % Clopper-Pearson 95% Interval
0.0 1501 8499 15.01% 14.32% — 15.73%
0.1 2477 7553 24.47% 23.36% — 25.32%
0.2 3114 6886 31.14% 30.23% — 32.06%
0.3 3528 6472 35.38% 34.34% — 36.23%
0.4 8032 11968 40.16% 39.48% — 40.84%
0.6 18697 21303 46.74% 46.25% — 47.23%
0.8 19288 20712 48.22% 47.73% — 48.71%
0.9 19870 20130 49.68% 49.18% — 50.17%
1.0 8214 11786 41.07% 40.39% — 41.76%
1.2 14639 25361 36.60% 36.13% — 37.07%
1.5 5868 14134 29.34% 28.71% — 29.97%
2.0 4707 25296 15.68% 15.27% — 16.10%

More visually, below is a plot of P0 win rate:

The results are quite interesting. A super non-aggressive AI (low aggressiveness) tends to fare badly against a standard (aggressiveness = 1.0) opponent. This result makes intuitive sense; the AI is trying too hard to clear the opponent’s board and ends up making unfavorable trades. At the opposite end of the spectrum, a super aggressive AI also doesn’t fare too well. This AI’s problem is that it tries too hard to go for the face, ignoring the opponent’s board, and ends up giving the opponents many opportunities to make very favorable trades. So, logically, it makes sense that the best AI should be somewhere between super-aggresive and super-controling.

The optimal aggressiveness in this setup seems to be around 0.9. To explain this, let’s think about the change in score when a play is made my the AI. When the AI decides to make a minion trade, what happens is that the AI gains score by killing an enemy minion, but also loses score by losing one of its own minion. If the two minions involved in the trade are of the same value (sum of health and attack weighted by their respective weights), the net change in score, assuming aggressiveness of 1, is 0. Thus, in this situation, the AI will think that the trade is not advantageous and will instead go for the face, which is at least score increasing. In summary, an aggressiveness = 1 AI will not make “equal value” trades. On the other hand, an AI with aggressiveness = 0.9 will usually make the equal value trade, because it scores its own minions slightly less than the enemy minions of equal value.

The fact that an aggressiveness = 0.9 AI fares better than an aggressiveness = 0.9 AI tell us that, when there is an equal value trade, it is usually better to make the trade than to go for the face. In fact, the results suggest that it is usually correct to go for a 2 for 1 trade as well, since an aggressiveness = 0.5 AI still performs better than an aggressiveness = 1.0 AI. Somewhat counterintuitive, but now wholly unbelievable.

3 thoughts on “HearthSim — Tuning the AI

  1. Shadi Isber

    Usually control has a higher W/L ratio, but you must also factor in average game duration. That’s why people use rush/aggro decks mainly. As long as the W/L ratio doesn’t suffer too much, you gain ranks faster simply by ending games in half the time.

    1. oyachai Post author

      Yep, great point. I starting tracking game duration as well.


Leave a Reply

Your email address will not be published. Required fields are marked *