Deep Reinforcement Learning Blind AI

This page contains the source code and model of our deep reinforcement learning blind AI (Blind AI), the details of which are given in this paper.

Installation:

Create and activate conda env:

git clone https://github.com/TeamFightingICE/BlindAI
cd BlindAI
conda env create -n ice -f environment.yml
conda activate ice

Boot DareFightingICE with the option --limithp 400 400 --grpc-auto --non-delay 0.
Run the train.py file to train. e.g python train.py --p2 MctsAi23i --encoder mel --id rnn_1_frame_256_mctsai23i --n_frame 1 --recurrent
Download the model from here and paste the folder named trained_model into the root directory.
Run the trained_ai/test.py to test the Blind AI. e.g. python trained_ai/test.py --p2 MctsAi23i --encoder mel

train.py is a file used to train Blind AI. Please run python train.py -h for our explanation of the parameters.
train_ai folder contains source code for the trained AI which is used as a sample AI for the AI track. It uses the FFT audio encoder and GRU. All weights are stored in trained_model.
visualize.py is used to visualize the learning curve and calculate the area under the learning curve.
analyze_fight_result.py is used to calculate the win ratio and average HP difference between Blind AI and MctsAi23i.

After finishing your sound design, please run the following command to train Blind AI: python train.py --p2 MctsAi23i --encoder fft --id {experiment_id} --n_frame 1 --recurrent, where you can decide experiment_id on your own
After training, a result file with the name result_fft_{experiment_id}_rnn.txt is created. Please run visualize.py as follows: python visualize.py --file result_fft_{experiment_id}_rnn.txt --title FFT. A plot will be shown and the area under the learning curve will be printed out.
Before testing the performance of the Blind AI against MctsAi23i, please remove all the files under log/point of DareFightingICE.
Please revise the path parameter of the Blind AI in line 16 of trained_ai/test.py to your trained model location.
Run python trained_ai/test.py --p2 MctsAi23i --encoder fft to begin testing.
After testing, please run python analyze_fight_result.py --path {path} where path is the location of log/point of DareFightingICE.
Both win ratio and average HP difference will be printed out.