Transcript Active Guidance for a Finless Rocket using Neurevolution

Active Guidance for a Finless
Rocket using Neurevolution
Gomez, F.J. and
Miikkulainen, R.
Content





Background
Enforced SubPopulations
Rocket Model
Experimental Results
Discussion
Background



Rockets used for scientific
research in upper atmosphere
Finless rockets can reach higher
than rockets with fins
Finless rockets are unstable, and
require active guidance
Background
Background

Highly complex environment:
–
drag is complex function of time, speed and altitude
Background


Neural networks can be used to model these
complexities without it knowing all the details
Use Enforced SupPopulations neurevolution to
learn an active guidance control system for the
RSX-2 rocket
Enforced SubPopulations

Extension of Symbiotic, Adaptive Neurevolution
(SANE)
–
–
Evolve individual neurons instead of entire network
Unlike SANE, use individual populations for each
neuron
Enforced SubPopulations:
Advantages and disadvantages



No loss of performance due to recombination
across different specialized neurons
Neurons are evaluated working together
Specialization causes diversity decline
–
–
–
Result: evolution gets stuck in local optimum
Remedy: burst mutation
When performance stagnates, add noise to the
populations according to some distribution
Enforced SubPopulations:
Algorithm (1/2)

Initialization:
–

Create a random population for each hidden unit
Evaluation:
–
–
–
Select a random neuron from each population
Determine fitness of resulting trial model
Repeat to accumulate fitness of individual neurons
over several trial models
Enforced SubPopulations:
Algorithm (2/2)

Recombination:
–
–

Neurons in top quartile average fitness are
recombined using 1-point crossover and mutation
Replace lowest half of neurons with recombined
neurons
Repeat evaluation and recombination until
performance of entire model is sufficient
Rocket Model



RSX-2 rocket is modeled in simulation package
A neural network determines the thrust of each
of the 4 thrusters between 90% and 100% of
maximum throttle
Fitness of a neural network is determined by
the altitude reached before burnout or failure
–
–
Failure occurs when a or b exceed 5 degrees
Networks with continuous 100% thrust get fitness 0
Rocket Model
Rocket Model



10 hidden units, 200 individuals per
subpopulation
Problem is too hard; neurons get stuck in a
local maximum at an early stage
Solution: incremental evolution
–
–
Perform ESP for a rocket with small fins
Use the output as a starting point for a finless rocket
Experimental Results
Experimental Results
Experimental Results


Controller generated without formal knowledge
of the system or correct behaviour
Future work
–
–
–
Remove a or b as input of the network
Change output of network to binary to simplify
rocket hardware
Incorporate noise and wind
Discussion