Dota2 Win Rate Prediction Model v1

Dota2 Win Rate Prediction Model is a model that we as VPGame's US research team build and trained in late 2016. The model is then used to build the project of AlphaMao, which predicts the win rate by inputing the line up of both teams. The model that supports the product is OUTDATED.

The training paused when we rebuilt and migrate our Dota2 Data System to GCP since Novmember 2017.

The project is now closed due to the release of Dota Plus. This is now a learning project for training and testing purposes.

DISCLAIMER: this code is an early research code. What this means is:

• It's not the exact same code when we use in production. But it shares the logic.
• It may not work reliably (or at all) on your system. Especially the libraries we were using are very old.
• We tried to do some brief clean up before uploading to Github. But the code quality and documentation are still quite lacking. The variable names may look random and we may not have a plan on improving that
• There are quite a few hacks made specifically for our systems and infrastructure.

Dependencies

We run the project using:

• Ubuntu 16.04 / Windows 10
• Python 3.5 / 3.6
• Keras 2.0.2
• Theano 0.9.0

With a newer version of Keras or Theano you may need to update some callings of functions like Merge since the way we are writing it may be deprecated.

Inside the Repo

We provided our training script along with our model. We also provide a test script so that you can test the result of your trained model

• Training: wrpt_train.py
• Model: wrpt_model.py
• Testing: wrpt_test.py

There are also some other supporting training/testing files.

• new_combo_2.vp: Win rate of two heroes when in the same team
• new_advantage_against.vp: Advantage Rating of one hero against another
• 2017-11-07_learning_data_1.vp: Training Data that includes 54339 matches of 2017-11-07 matches. On average, there are 600,000 valid matches in one day.
• train6_v707_new_temp.weights: Sample training result using the provided training data

To Train

cd dir_to_project/

python wrpt_train.py

or if you have python2 as default python

python3 wrpt_train.py

When an epoch of training is finished, you should be able to see a screen like this:

54339/54339 [==============================] - 73s - loss: 0.6745 - hid_layer_loss: 0.6733 - another_output_loss: 0.0024 - hid_layer_acc: 0.5820 - another_output_acc: 0.2177 - val_loss: 0.6889 - val_hid_layer_loss: 0.6883 - val_another_output_loss: 0.0012 - val_hid_layer_acc: 0.5091 - val_another_output_acc: 0.2000

Where hid_layer_acc is the overall accuracy. In this case, it's 0.5820. As you can see we trained this model using 54339 matches of 2017 Nov. 07th.

To Test

python wrpt_test.py #hero_id_1 #hero_id_2 #hero_id_3 #hero_id_4 #hero_id_5 #hero_id_6 #hero_id_7 #hero_id_8 #hero_id_9 #hero_id_10

where hero_id_# (1-5) are the heroes of Radiant and hero_id_# (6-10) are the heroes of Dire.

It takes time to load the weight into the model and to build the Neural Network. After it's done you should be able to see a result like this:

[ 0.351208    0.64879197]

Where the first value is the win rate of Dire and the second is the win rate of Radiant. Keep in mind that the order may look reversed to you. This is because in Dota 2 Radiant is considered as 1.

In this case, we simply input

python wrpt_test.py 1 2 3 4 5 6 7 8 9 10

which means Radiant has a line up of

• Antimage
• Axe
• Bane
• Bloodseeker
• Crystal Maiden

and Dire has a line up of

• Drow Ranger
• Earthshaker
• Juggernaut
• Mirana
• Morphling

The win rate of Radiant (before playing the game) is 64.9%

We also offered overwrite code inside wrpt_test.py (Line #24, #25):

# rad_squad = [1,2,3,4,5]
# dire_squad = [6,7,8,9,120]

You can uncomment these two lines and set the hero_ids there as well.

Hero ID look up

Please don't abuse the API Key, if the api key is banned we are not going to update it.

Understand the Accuracy

(WIP)

Some Improvements we made but not shown in the code

• The model doesn't know that the order of a team doesn't affect the result. For example, if you input 1 2 3 4 5 and 5 4 3 2 1 for Radiant, the result will be different (but really close). What we did is for each match input, [h1 h2 h3 h4 h5], we created a group of "cloned" matches with same heroes, but in different order. After a large scale of training, the model will learn that the order doesn't matter. Another benifit of doing this is this will create a lot training data and help your model to learn better with limited amount of match data.

Common Questions

Where to get more training data?

We rebuilt our Dota2 Data System and now it's under VArena. It's still in Beta and currently it cannot support the sufficient data. It will be online soon.

On the other hand, if you don't mind old data, we are also planning to host our former training data (includes 100,000,000+ matches before 2017 July) this summer. For now you can download these 599868 matches data to play around. This should be enough to make your model to reach a 58+% overall accuracy.

Discussion

If you have anything you'd like to discuss. Feel free to shoot an email to [email protected] or [email protected].

If you would like to join us for more fancinating game related AI project, check this document