Strategy optimization in a dynamical financial analysis environment through evolutionary reinforcement learning

This thesis develops a reinforcement learning framework to solve insurance control problems. A Dynamic Financial Analysis model is formulated to represent the environment in which a non-life insurance company operates. Based on the modelled environment, a multi-objective stochastic control problem is formalized by defining the company’s control variables and target quantities to optimize. To avoid a modelling bottleneck induced by analytic techniques, two computational methods, neural networks and symbolic regression, have been adopted to approximate candidate strategies. Depending on the approximation method, strategies are represented by a specific set of parameters. Therefore, the search for optimal strategies boils down to the search for an optimal configuration of such parameters. To this end, an evolutionary inspired search algorithm has been adopted and compared to a Uniform Monte Carlo Search. Numerical results show that the proposed framework can find optimal strategies regardless of the underlying insurance model complexity or number of control variables.