atremblay / agent-cribbage

agent-cribbage

There are 4 jobs that you can call in this module:

1- Play against an agent or make play 2 agents agaisnt each other (requires a .yaml configuration)
2- Preheat a model (Only works for The Show phase model).
3- Train agents models (requires a .yaml configuration)
4- Evaluate models in a given directory (requires a directory containing one or many configurations to evaluate, and the path where the models are located).

For you to play against our best agent:

python -m agent-cribbage Play --agent_yaml ./agent-cribbage/Configurations/Play_agent_vs_human.yaml

For you to play against the simple deterministic agent:

python -m agent-cribbage Play --agent_yaml ./agent-cribbage/Configurations/Play_deterministic_vs_human.yaml

To make play multiple agents against each other multiple games:

python -m agent-cribbage Play --agent_yaml ./agent-cribbage/Configurations/Play_agent_vs_random.yaml --number_games 1000

To preheat the model for phase 0 (The Show):

python -m agent-cribbage Preheat --model conv --epochs 3

To train a model giving a configuration:

python -m agent-cribbage Train --epochs 500 --number_games 200 --agent_yaml ./agent-cribbage/Configurations/Train_agent_ConvLstm_Dealer+split_discarded_MC.yaml --dataepochs2keep 5

To evaluate a directory containing multiple models:

python -m agent-cribbage Evaluate --number_games 5000 --config_dir ./agent-cribbage/Evaluate/Train_agent_ConvLstm_Dealer+split_discarded --model_dir ./Train_agent_ConvLstm_Dealer+split_discarded_MC

agent-cribbage's Issues

Trainer that use BackupDiagram instance to train network

See section 7.6 of Sutton

Create script for agent VS human

Use Job class

Generic Backup diagram class for algo definition

CNN approach for poker

https://www.aaai.org/ocs/index.php/AAAI/AAAI16/paper/view/12172/11606

Create sanity test for learned agent strategy

Add index of choice in data dump

Use TensorBoardX

Minimum implementation of render and the other mandatory method of gym.

Minimal implementation for gym.

One game per pickle file - fix

investigate deepstack (Poker AI)

http://science.sciencemag.org/content/356/6337/508.full

Sequencer: alternate between data generation and training

Use latest model to generate new trajectories. Then, use these trajectories to train new value functions.
We could train the model with old data using a decaying learning rate(old data small lr while new data larger lr) up to a threshold.

Investigate this hand evaluation code

https://github.com/erbaker/Cribalyzer