Variable reinforcement: random rewards are harder to quit than reliable ones