Evolutionary Neural Network Outperforms Humans and PID Controller in Single-Engine Glide Distance Study
Why It MattersThe findings point toward autonomous glide-optimization systems as a potential safety tool for general aviation piston aircraft facing engine failure scenarios.
A study comparing control approaches for extending glide distance after single-engine failure found that automatic controllers outperformed untrained human pilots. The research used a Python-based two-dimensional flight simulator modelled on the Cessna 152, validated to within 0.4% of the aircraft's handbook glide performance of 1.6 NM per 1,000 ft, using Runge-Kutta integration and altitude-varying air density.
Three approaches were tested: seven untrained human participants using a slider interface, a PID controller holding a constant flight-path angle, and an evolutionary neural network (ENN) whose weights were set by a genetic algorithm and which was free to determine its own elevator inputs. At the starting condition used for the human trials, the PID controller glided 17.03% further than the average human result, while the neural controller glided 17.77% further.
The two automatic controllers were then tested across 400 combinations of starting altitude and airspeed, 390 of which were never seen during training. Across those unseen conditions, the neural controller outperformed the PID controller by 35.73% on average. The study noted limitations including the simulator's two-dimensional scope, absence of wind or turbulence, coverage of only one aircraft type, and reliance on trials with seven untrained participants using a simplified interface. It also observed that holding a fixed best-glide speed by hand is difficult and that the optimal glide speed may not be a single fixed value under varying conditions.
The Australian Transport Safety Bureau reports a piston-engine failure rate of roughly 1 to 6 per 10,000 flight hours, and engine problems remain a recurring cause of general aviation accidents in the United States.
