RL-testing-ground

Evaluation of different RL decision methods against the K-armed bandits test

Running the code

Needs Matplotlib and numpy. Values for the number of episodes, arms (k), and iterations can be changed in lines 4,5, and 6 of the file.

Current Decision methods

• Elipson greedy. A threshold is given - this theshold represents the probability that the desicion method will either explore (do a random action) or exploit (do the highest valued action from current data)
• Elipson decay. Like elipson greedy but the threshold decays exponentially with experience
• Upper confidence bound. Chooses the maximal action, but adding "optimistic" value based on how much experience it has with each arm (more uncertain, more optimistic) in a logarithmic fashion.