An Evaluation of Stochastic Model-Dependent and Model-Independent Glider Flight Management

This paper develops and compares two stochastic strategies that manage glider flight in a dynamic environment, to tackle an open problem concerning the necessity, applicability, complexity, and validity of using an environment model when making flight management decisions. Strategy performance is compared on two surveillance and pursuit tasks that require optimal control: the maximization of expected range while maintaining altitude within given limits, and the maximization of expected range while following a moving ground vehicle within a prescribed distance and maintaining altitude within given limits. Both tasks involve flight management decisions of glider airspeed and the time-to-spend climbing in a randomly encountered thermal. The first strategy is stochastic drift counteraction optimal control (SDCOC), a dynamic programming method that relies on a model. Here, SDCOC uses a computationally inexpensive yet accurate environment model that reflects transition probabilities between updrafts and downdrafts as well as thermal locations and strengths based on existing glider flight data. The second strategy is selective evolutionary generation, a model-free Markov chain Monte Carlo method that is not subject to the curse of dimensionality, efficiently searches for an optimal control in an online and tunable fashion, and easily adapts to a dynamic environment. However, this strategy requires a tolerance for learning and decision-space exploration. Both strategies perform satisfactorily, and showcase an environment/decision-space exploitation-exploration tradeoff.