Created by W.Langdon from gp-bibliography.bib Revision:1.5700
In my dissertation research and related work, I evolved strategies by which an aircraft could evade anti-aircraft missiles. The approach I took to fitness evaluation was to simulate an encounter between a missile (using proportional navigation) and an aircraft (controlled by stick and throttle commands issued by a control program). The simulation ran at 50 Hz (typical of aircraft flight control computers) Fitness was equated to aircraft survivability. The training population consisted of missiles launched from numerous potentially lethal positions. Aggregate program fitness reflected aircraft survivability against each missile in the training population (i.e., program X survived 25 out of 50 missiles in the training population; etc.). Best-of-run programs optimised survivability against the training population, and were subsequently tested against a large, representative test population of missiles to see how well the evolved solutions generalised.
The problem with using simulation to evaluate fitness is that one has to execute each program from the evolved program population over N simulated time intervals, just to determine fitness against a single training case. (For my missile problem, typical simulated encounters lasted 20 seconds, thus entailing 1000 program executions PER FITNESS CASE.) So, we're talking about 2-3 orders of magnitude more computation than is typical for GP fitness evaluation. For the CPUs available to me, it was not uncommon for a run to take several days to complete. BUT the best-of-run program was an embedded real-time controller that executed specific aircraft manoeuvres (and, later on, deployed specific countermeasures) to optimize aircraft survivability. What makes that significant is the fact that, for the general missile countermeasures optimization problem under conditions of uncertainty about missile type and/or state, NO ANALYTICAL SOLUTION METHODOLOGY currently exists. I believe that by combining genetic programming with sophisticated simulators, we will be able to optimise programs that solve a wide range of control problems for which analytical solutions are difficult or impossible to identify. I'd like to see GP research move away from toy problems and onward to complex real-world applications, and I think this approach could help further that process.
Regards to all.
OCLC Work Id: 1864647562",
Genetic Programming entries for Frank William Moore