☰
We've trained AI models to predict the winrate of each game only using champion selections
The linear model uses a combination of champion winrates, synergies, and counters. These values are calculated by taking games with certain champion combinations, and calculating the change in winrate, and summing these values. This means it is fully explainable, and we can see how the model makes its predictions
This prediction is corrected via a black box AI model, we keep the workings of this proprietary. This model is usually more accurate, and is not sensitive to lane ordering
We measure model accuracy with Expected Calibration Error (ECE). This is the difference between the predicted winrate and the actual winrate of that group. The lower the better.