Armed Bandits

• In AI and decision theory, armed bandits are gambling devices. They're like slot machines.
• What is typically done with a one armed bandit is to show how likely you are to get a positive return, with the rest of the options being 0.
• For instance, you have 10% chance of winning £100 and 90% chance of losing your bet.
• Often people are asked what they'd bet to play this bandit.
• The key is that the experimentor can set the probabilities, or arrange payouts explicitly.
• You can use this to show how people don't choose rationally (Kahnemann)
• There is also a lot of work with multi armed bandits (K or N).
• Part of the question here is, if you can play over and over again, how do you discover what to play.
• If you pick the right bandit, you can maximise your expected outcome.
• This involves the exploration exploitation dilemma.
• Moreover, the probabilities can change over time.