Game-playing ANNs that use a stumbler-strategist architecture. Stumblers learn to map states to actions. Strategists study stumblers, learning to find the structure within the states.
We focus on co-operative AI. By sharing strategies our ANNs try and help people, not replace them.
Based on ideas from: Muyesser, N.A., Dunovan, K. & Verstynen, T., 2018. Learning model-based strategies in simple environments with hierarchical q-networks. , pp.1โ29. Available at: http://arxiv.org/abs/1801.06689.
We come at ANN design as scientists and humanists. Meaning we take three strong philosophical stances:
- ANNs should help people, not replace them.
- To win in the long-term ANNs and science must create a close virtuous cycle of improvement. Principled science--psychology, biology, neuroscience--should directly inform ANN design. ANN results should directly inform science.
- ANNs must help us understand our problems better. This means an ANN must always be able to explain itself to a person.
- python3
- a standard Anaconda install
- pytorch
- fire (https://github.com/google/python-fire)
- tensorflow
- tensorboard
- tensorboardX
Install instructions: https://github.com/lanpa/tensorboard-pytorch
- From the command line run
git clone https://github.com/CoAxLab/azad.git
- Then from the top-level
azad
directory run,pip install .
for a normal install orpip install -e .
if you are going to be editing the code.
In Ian Banks delightful book, The Player of Games, master game player Jernau Morat Gurgeh travels to the planet Azad to play the game Azad. Though it takes a lifetime to master, Azad is as much a statement of philosophy as it is a game of winning and strategy. In the book, Gurgeh comes from an alien culture so his philosophy and play is quite different than his opponents.