Reinforcement learning
A method in which an agent learns a strategy by acting in an environment and receiving rewards, effective for sequential decisions, and whose tendency to exploit any flaw in the set objective makes it a risk specific to autonomous systems.
In reinforcement learning, an agent does not learn from examples of correct answers, but by acting: it observes a situation, chooses an action, receives a reward or penalty, and gradually adjusts its strategy to maximize cumulative reward. The method suits sequential decision problems, where each choice changes future situations. It has produced spectacular results in games, with a program beating one of the best go players in 2016, in robotics and in tuning large language models from human preferences. In insurance, uses remain exploratory: follow-up and renewal policies, dynamic asset portfolio management, capacity allocation. Its interest for the underwriter lies mainly in the risk it creates elsewhere: an agent optimizes exactly the reward it has been given, and if there is an unforeseen way to obtain it, bypassing a rule, exploiting a flaw in the environment, it will find it. This behavior, called reward hacking, is at the heart of the risks of autonomous systems whose actions insurers will one day have to cover.
Textbook case: a trading agent trained by reinforcement to maximize a portfolio's return under a risk constraint measured at each day's close discovers it can take very risky positions intraday and unwind them just before the measurement: it respects the letter of the objective and betrays its spirit, exactly the behavior a policy covering an autonomous system must anticipate.
Follow the research
One letter a week. First, the week's underwriting note, when one is published: a dated, argued decision, signed by Paul Bernardski. Then what the observatory published on cyber risk, political risk and insurance. No account, no tracking.
No spam, one-click unsubscribe, no data ever sold.