- What does loltorch do?
- Requirements
- Installation
- Trying loltorch out
- Training your own networks
- How does loltorch work?
- Neural Network Details
- Caveats
- Disclosure
loltorch pulls match data from the League of Legends API and runs the data through a neural network which tries to make the Masteries, Runes, Spells, and Items look more like what the winning team used and less like the losing team. This predicts what the optimal winning build is for the combination of 1) Champion 2) Role 3) Lane 4) Region and 5) League of Legends version. Even if given a combination of those attributes not seen in the data it will make a prediction as to what an optimal build might look like. For example, if you wanted to know what a support character like Blitzcrank might need in order to be an effective jungler, this project aims to do just that.
This project depends on:
- Install Torch.
- Install the following additional Lua libraries:
luarocks install lualol
luarocks install nn
luarocks install nngraph
luarocks install tds
luarocks install threadsNote if you have any trouble are trying to install loltorch on OSX from within torch's LuaRocks installation you may need to use following command if you get an error about not having OpenSSL installed.
> luarocks install lualol OPENSSL_DIR=/usr/local/opt/openssl/
outdir=pretrained version=6.13.1 ./bin/getstaticdataNow that you have the static data downloaded you can use the following command
modeldir=pretrained ./bin/sample mlp384 -datadir pretrainedto get a preset list of examples comparing various roles for a subset of champions. Using the following command
modeldir=pretrained ./bin/sample_odd mlp384 -datadir pretrainedwill give some examples of weird combinations such as having Blitzcrank as an adc.
For trying out your own combinations simply use the command:
th sample.lua -helpto see a list of options.
In order to train your own network on new data you need to first get the match data and process it:./bin/getstaticdata
./bin/getallmatches
./bin/processmatchesThen simply train a model using the data. Use the following command to get a list of training options (and their defaults):
./bin/trainbuildNote that when actually training, the first parameter to the trainbuild script
needs to be the model type, so the minimal command to train the gated384 model
would be:
./bin/trainbuild gated384loltorch first pulls all the ranked match data from the League of Legends API for the Master and Challenger leagues then takes all the features of the match (Champion, Role, Runes, Spells, etc) for each participant and puts them into a corpus.txt which is then run through word2vec in order to generate vectors which makes closely related features more similar to each other. Then when training the neural network tries to make the cosine similarity of the features of the winning team large and those of the lossing team look smaller. This has the effect of making the final champion builds that this project predicts to be what is optimal for winning a match.
When actually sampling predictions from the neural network it generates an output of vectors that are then compared to the feature vectors to see which ones are closest in similarity. This similarity value is then able to be output (along with filtering it based on a threshold), such that you can see how confident the neural network is in predicting that particular feature.
A number of architectures were tried in regards to what would be an effective neural network. The final neural network models that were chosen for inclusion are visible in the 'models.lua' file. Through the use of cross validation the best neural network results were obtained by using the 'gated768' model which uses a separate MLP (multi-layer perceptron) per input feature and then combines them based on a separate neural network which is the 'gate'.The results of the final k-fold cross validation are listed in the pretrained/README.md file, showing the parameters used for training and cross validation scores of each network tested.
Since I only have a 2013 model Macbook Pro with a Core i7 I only implemented the neural network using the threading library in Torch7. While the laptop I have has a dedicated GPU, in initial testing it turned out to be much slower than running threads on the Core i7, so no CUDA support is built into the training scripts.Since the Core i7 is very slow in generating some of these results there may be much better networks (that have many more parameters) than the ones I chose. For example, the 'gated768' model has better cross validation results than the 'gated384' model, but since it took over three days to obtain the cross validation results on only a subset of the total data there was only so much time I was willing to commit to determining better models. If someone is so inclined and has powerful enough GPU(s) they could add support for CUDA and try out larger models to see if they produce better results.
The approach of using word2vec to generate the vectors was a decision based on performance. Initially an approach using Torch7 was used, but this turned out to be much slower. For that reason word2vec with a text corpus was chosen as a close analogue to generating vectors based on the input features using Torch7.
loltorch isn't endorsed by Riot Games and doesn't reflect the views or opinions of Riot Games or anyone officially involved in producing or managing League of Legends. League of Legends and Riot Games are trademarks or registered trademarks of Riot Games, Inc. League of Legends © Riot Games, Inc.