og-nn-sim

An attempt to train the battle engine with neural networks.

Requirements

• C++20 compiler (for dataset-gen; currently only Clang manages to compile it)
• CMake >= 3.12
• Python >= 3.7
• Tensorflow 2
• Pandas

Instructions

Building dataset generator

cmake -S ./dataset-gen -B build -DCMAKE_BUILD_TYPE=Release -DDATASET_GEN_ENABLE_ARCH_NATIVE_OPT=On
cmake --build build

Generating a dataset

This will generate a dataset and save it into dataset file.

./build/dataset-gen --dataset-size 1000000 --max-ships 10 --out dataset

Check also ./build/dataset-gen --help

Training

Now, you can train the network on the generated dataset. This will save the trained model info model directory and the normalization scales into scales file.

./train.py

Run battle example

./battle.py

Optimize model

Optionally, you can try to find a better number of layers or units in the model for your dataset:

tensorboard --logdir logs &
./optimize-model.py