This project uses machine learning (Logistic regression and Neural Network algorithm) to calcuate the winning chance for Dota2 matches. The workflow is as follows:
-
The
matchfiner.pymust be used first to retrieve all of the matches for the player. At this stage, player's Dota2 account ID is hard-coded in this script. It uses Open Dota API. As we are using the free version of the Open Dota, due to the API call limit, this script may take some hours to retrieve all of the matches. The final result will be a large JSON text file namedmatches.txt. -
The
featurize.pythen should run to open thematches.txtand extract the required features of the matches. For this project, the following features of matches are used:- Radiant score - Dire score
- Radiant average total gold - Dire average total gold
- Radiant average total XP - Dire average total XP
The positive value of the above features indicates the Radiant's team dominance. The result of this code is the
dataset.txtwhich has four columns, three of them are the aforementioned features and the last one is the result of each match. 1 indicates Radiant win and 0 means Dire win. -
Then for the last step, we have run
logisticReg.mwhich is an Octave code to run our Logistic Regression algorithm and find the best fit for the provided dataset. At this time, the outputs of this code are:- A plot that shows the learning curve of the algorithm.
- A plot that illustrates the dataset using x for Radiant win case and o for Dire win matches. Because there are three features, in order to plot a 2-D diagram, the score difference was used as the x-axis and average of gold and XP difference was used for the y-axis. Also, this plot includes a simple contour for seperating the win/lose case based on the parameter theta found by the logistic regression.
- The accuracy of the algorithm for training data and test data.
-
Alternatively, the
neuralnetwork.mcan be used for the same purpose. It uses a 5-layer architecture (3 hidden layers each has 5 units including bias units).