Avalanche
Snowball
From avalanche/snowball directory, run:
python run.py --adversary_percent=0.19 --adversary_strategy EQUAL_SPLIT experiment
Observe 4 plots that show nodes color assignment and their confidence in that color. If at the start already one color is prevalent - restart.
Example of snowball being very confident in both colors for 1000 iterations:
See adversary.py for more adversary strategies.
Supervised framework pipeline
From avalanche/snowball directory:
Create target dataset
python run.py --adversary_strategy TRY_BALANCE learning --create_dataset
Train model using this dataset
python run.py --adversary_strategy TRY_BALANCE learning --train
Test model performance on protocol simulation
python run.py --adversary_strategy RL --net_name supervised-0 experiment --iterations_per_frame 100
Consider using low number of --iterations_per_frame or --no_plt.
Evaluating the network at every adversary step is expensive.
Train with reinforcement learning
Current training algorithm is A3C. To start training
python run.py rl
Credits
This code is copied from a private branch, and is initially built by Marcelo Fornet.