Bayesian reinforcement learning to optimize paid ancillary revenue in the airline industry

To optimize the pricing of paid ancillary seats, we adopt a revenue management approach that optimizes over the capacity of these seats while accounting for unknown underlying model parameters. We test various models against a simulation model to assess the performance against wide-ranging input parameters. We demonstrate that using a Bayesian exponential demand model to describe the relationship between price and seats sold, combined with a Bayesian reinforcement learning approach to estimate its parameters, outperforms other approaches. By using a relatively simple demand model with a limited number of parameters, updating in a Bayesian manner, and in one step estimating demand parameters to directly use for price optimization, the model is quickly able to perform well across a wide range of demand scenarios.