Skip to content

Latest commit

 

History

History
64 lines (47 loc) · 2.71 KB

README.md

File metadata and controls

64 lines (47 loc) · 2.71 KB
Raw

Learn The Spire

Neural Network for choosing between cards in Slay the Spire using Tensorflow.

Requirements

python3 and pip3

pip3 install -r requirements.txt

Should be able to locate runs on Windows and macOS.

Main Idea

  • Scrapes local run data into IRONCLAD_TRAINING_DATA, THE_SILENT_TRAINING_DATA, and DEFECT_TRAINING_DATA, using get_training_data.py, and encoding cards into 1 x 234 vectors with card_name_to_vector.py.
python3 get_training_data.py
2018-11-19 19:32:55,940 - learnthespire - INFO - Found 37 IRONCLAD runs
2018-11-19 19:33:07,705 - learnthespire - INFO - Found 1 DEFECT runs
2018-11-19 19:33:07,895 - learnthespire - INFO - Found 4 THE_SILENT runs
  • Train a feed-forward neural network with our past run data with train_model.py, using the supplied training data file and save it to the disk.
python3 train_model.py IRONCLAD_TRAINING_DATA
2018-11-19 19:33:30,393 - learnthespire - INFO - Epoch 1 completed out of 100 loss: 719.8838911056519
2018-11-19 19:33:30,444 - learnthespire - INFO - Epoch 2 completed out of 100 loss: 551.330732345581
2018-11-19 19:33:30,496 - learnthespire - INFO - Epoch 3 completed out of 100 loss: 456.9212851524353
...
2018-11-19 19:33:35,955 - learnthespire - INFO - Epoch 98 completed out of 100 loss: 21.039612337946892
2018-11-19 19:33:36,007 - learnthespire - INFO - Epoch 99 completed out of 100 loss: 20.799650326371193
2018-11-19 19:33:36,062 - learnthespire - INFO - Epoch 100 completed out of 100 loss: 20.55322101712227
2018-11-19 19:33:36,136 - learnthespire - INFO - Accuracy: 79.49526906013489%
2018-11-19 19:33:36,225 - learnthespire - INFO - Model saved to ./IRONCLAD.model

After a hundred epochs, which takes around 5 seconds with this small dataset, we can train a neural network that chooses the same as a human player around 80% of the time.

At around 500 epochs, it can pick with what seems to be 100% accuracy on the dataset, but in reality it's probably much lower because the testing and training sets are the same and there's a small dataset so there could be overfitting.

  • Load the model and use it to tell us which card to pick.
python3 load_model.py IRONCLAD.model
Card 1: barricade
Card 2: limit break
Card 3: offering
The Neural Network chooses: Limit Break
Card 1: shrug it off
Card 2: true grit
Card 3: flex
The Neural Network chooses: True Grit

  • Go back and evolve this to take more inputs like our existing deck, and maybe even support choosing choices at random events and route choices.

  • Fully automated AI Slay the Spire player.

Credit

The structure of the neural network is heavily based on the one created in sentdex's series on neural networks.