Herman Persson: Insurance Premium Control with Deep Double Q-learning

Abstract.

In this project, we investigate how deep reinforcement learning, specifically a Double Deep Q-Network, can be applied to solve a stochastic control problem in insurance premium control. By building on existing actuarial models, we construct a simulated insurance environment in which the insurer aims to maximize long-term surplus while minimizing the risk of default. By formulating the premium control problem as a Markov Decision Process, we show how a reinforcement learning agent can learn effective pricing strategies from simulation. Our results indicate that the agent can converge to a policy which increases the surplus, but there exists large issues with training stability. We discuss how training efficiency and stability can be improved through statespace reduction and model simplification. Furthermore, we outline potential extensions to more complex multi-product and investment scenarios. This work highlights the promise of deep reinforcement learning as a tool for modern actuarial decision-making, particularly in dynamic, data-rich environments where traditional methods may fall short.