Dota 2 is a popular video game played competitively as an eSport. An important part of every Dota 2 game is the draft: the phase during which each team of five players picks what heroes (characters) they want to play with.

This may sound simple, but drafting well is deceptively complicated. Each Dota hero has a unique set of abilities, and the specific interaction of one hero’s abilities with those of its teammates might allow for an awesome combination play. Similarly, another ability might completely mitigate the destructive potential of the enemy team. Among professional Dota players, drafting is complicated artform that decides the outcomes of million-dollar games.

Draftnet learns how to draft well by analyzing thousands of winning drafts by professional Dota players. These games are read from the OpenDota API, and then used to train a feed-forward neural network that makes drafting decisions.

You can find an easy-to-use web interface for Draftnet at http://draftnet.herokuapp.com/. The website is designed so that you can have it open in a secondary monitor or tab while you are picking a Dota 2 draft. Because of temporary limitations on the OpenDota API, the network currently only predicts picks and bans for Captain's Mode games.

Our Python implementation of Draftnet requires the Tensorflow machine learning library. After you have installed Tensorflow and cloned our repository, you can train the prediction network with draftnet.py. For example:

python draftnet.py --train train/pro-7.00.json --test test/pro-7.00.json --save models/pick/custom.ckpt

Train and test files contain game data fetched from the OpenDota API. Type python draftnet.py --help to view a full list of run options.

Draftnet uses a "bag-of-heroes" architecture inspired by feed-forward neural networks like Word2Vec that learn vector semantics for words. The network takes as input four k-hot vectors representing sets of heroes (the picks and bans on each team) and one bit that encodes whether the next draft action is a pick or ban. These inputs feed into a hidden layer with ~50 bits, which in turn feeds into a final output layer with 113 bits (corresponding to the number of heroes in Dota). The softmax of this output layer is interpretted as a probability distribution for the next pick or ban. The network is trained via gradient descent. The loss function in this optimization is the cross entropy between the network's predicted distribution and the one-hot distribution encoding the actual pick or ban.

In addition to predicting hero picks, our neural network also produces embeddings of each hero in Dota. Intuitively, an embedding is a vector representation of a hero that captures all the important features that that hero has. Since the network stores these embeddings as vectors, we can use them to visualize the relationships between heroes. You can use tsne.py to produce a plot of these hero vectors in 2D space. For example:

python tsne.py --model models/pick/pro-7.00.ckpt