For a Reinforcement Learning class, I worked on a few popular algorithms for the bandit problem. Among them were :
- Epsilon-greedy bandit
- BESA
- Softmax
- UCB1
- Thompson sampling
- KL-UCB
Bandits are implemented in agent.py
For the purpose of the class, each agent was tested on a specific configuration : 1000 rounds for 2000 agents in parallel :
python main.py --niter 1000 --batch 2000
Use python main.py -h to know more.