Simple Models of Discrete Choice and Their Performance in Bandit Experiments