Links:
- Fix MRAS-Categorical-Subset reproducibility
- Tune MRAS-Dirchlet-Subset
Use python 3.6
.
pip install -r requirements.txt
To run:
export PYTHONPATH=.
python sim.py
- UCB
- Thomson Sampling
- Aysm-UCB
- KL-UCB (Needs to be sped up)
We try the following parameter distributions:
- Categorical
- Dirchlet
- Gaussian
We also experiment with the following:
- Increasing function
H
. - Exploitation param
lambda
. - Simulation allocation
M_k
. - Population size
N_o
- Add under
bandits
folder. - Add a unit test to verify its working.
- Import it under sim.py
We use sacred for configs and capturing arguments. Read more here. Modify base-config.yaml for more games.
If you need very different args, create a new config file and run as:
python sim.py with configs/new-config.yaml
- Is regret being computed correctly? Right now we are accumulating (best_mean - reward). This could be negative, but averaged over experiments is positive.
- Is the UCB implementation correct?