Bandit Observation Engine: simulating context-free bandits.
First, create a YAML file similar to egreedy-basics.yaml.
To see run options, including param names, type:
python -m boe.experimenting --help
Note that contents of the py directory should be included in PYTHONPATH.
- Delayed reward (initial delay)
- Periodic update (periodic delay)
- Periodic snapshot of bandit state
- HTML report on experiment
There are four reported metrics, which are saved at each snapshot:
- Probability of selection per arm
- Average reward per arm
- Cumulative reward per arm
- Global cumulative reward
There is an option of reporting the Upper Confidence Bound (UCB) along with the average reward per arm. The parameter report-ucb defines this behavior.
When UCB is reported, there is an option to discount the value, linearly. This is strictly for plotting purposes. UCB is relative measure, and as such, should be used to estimate the confidence of knowledge of one arm over other arms. Not as an absolute measure.
The computed UCB at each snapshot is stored as is.
Use python3.