Sam Stites

Intrinsic Fear [paper]

November 7, 2016

Cleverly posted before the US election. This paper talks about the introduction of “Intrinsic Fear” for DRL algorithms.

The idea here is that DRL is being used in more and more life-or-death scenarios, but function approximation will make a RL algorithm to forget an infrequent occurence and repeat the mistake — turning into the Sisyphean Curse.

They basically introduce a supervised “danger model” alongside their DQN implementation which models the analogy to “catastrophic failure” — I’m thinking something to the effect of “black swans.” This danger model then becomes an extra source of reward (or, rather, cost). Check it out at: “Combating Reinforcement Learning’s Sisyphean Curse with Intrinsic Fear