In this experimental study, we use a data-driven machine learning framework based on genetic programing (GP) to discover model-free control laws (individuals) for suppressing self-excited thermoacoustic oscillations in a prototypical laminar combustor. This GP framework relies on an evolutionary algorithm to make decisions based on natural selection. Starting from an initial generation of individuals, we rank their performance based on a cost function that accounts for the trade-off between the state cost (thermoacoustic amplitude) and the input cost (actuator power). We then breed subsequent generations of individuals via a tournament in which the direct forwarding of elite individuals occurs alongside genetic operations such as mutation, replication, and crossover. We implement this GP control framework in both closed-loop and open-loop forms, followed by benchmarking against conventional open-loop control based on time-periodic forcing. We find that while all three control strategies can achieve similarly large reductions in thermoacoustic amplitude, GP closed-loop control consumes the least actuator power, making it the most efficient. It achieves this efficiency by learning an actuation mechanism that exploits the strong heat-release-rate amplification of the open flame at its preferred mode, even though the GP algorithm has never seen the open flame itself. This study demonstrates the feasibility of using GP to discover new and more efficient model-free individuals for suppressing self-excited thermoacoustic oscillations, providing a promising approach to data-driven feedback control of combustion devices.