Evo star2012 Robot Base Disturbance Optimization with Compact Differential Evolution Light
1. Robot Base Disturbance Optimization with
Compact Differential Evolution Light
Giovanni Iacca, Fabio Caraffini,
Ferrante Neri, Ernesto Mininno
University of Jyväskylä
Faculty of Information Technology
evo* 2012 - Málaga, Spain
April 13th
2012
3. #giovanni_iacca @evostar_2012_málaga
Memory Issues in Computational Intelligence
Computational Intelligence Optimization is good!
Why, Master?
Because its methods are efficient and robust!
Only that? What else, Master?
In addition to that, they often don't require any information
about the problem at hand, my son.
Oh, cool! But is there any drawback?
Well, actually most of them use plenty of hardware resources
and are computationally expensive.
Mmm...
5. #giovanni_iacca @evostar_2012_málaga
Memory Issues in Computational Intelligence
Single Solution Algorithms
- Simulated Annealing (SA)
- Single Particle Optimization
- Some Local Search Algorithms (e.g. Hooke-Jeeves)
Compact Optimization
- compact Genetic Algorithm (cGA)
- compact Differential Evolution (cDE)
6. #giovanni_iacca @evostar_2012_málaga
Estimation of Distribution Algorithm (EDA)
Does not use a population of individuals → low memory footprint
Makes use of a statistic model of the population:
Multivariate Gaussian with decision variables normalized in [-1,1]
Convergence: shrinkage of the Gaussian over the best (“elite”)
Sampling introduces beneficial randomness
compact Differential Evolution (cDE)
8. #giovanni_iacca @evostar_2012_málaga
DE can be straightforwardly
encoded into a compact algorithm
without losing the basic working
principles (instead of cGA)
Survivor selection scheme
(one-to-one spawning logic)
Persistent vs non-persistent elitism
Parameters: F, Cr, Np
compact Differential Evolution (cDE)
Original implementation
(cDE rand/1/bin)
9. #giovanni_iacca @evostar_2012_málaga
Mutation Light: only one solution is sampled (instead of 3)
→ →
compact Differential Evolution light (cDElight)
Statistic independence
Truncated Gaussian ~ Gaussian
Under the following
assumptions/approximations:
Limited computational overhead!
10. #giovanni_iacca @evostar_2012_málaga
Exponential Crossover Light: ad hoc xover exp instead of xover bin
compact Differential Evolution light (cDElight)
Exponential Crossover (standard) Exponential Crossover Light
only one
random number
multiple
random numbers
→
11. #giovanni_iacca @evostar_2012_málaga
Computational Overhead
(vs cDE)
compact Differential Evolution light (cDElight)
Computational Overhead
(vs different algorithms)
~1/3 computational
overhead - O(n)
Tests performed on a single core Pentium 2.8 GHz PC, with the sphere function (Java implementation) for
different dimensions (2:100). The algorithm overhead is computed on 10000 fitness evaluations.
memory-saving
population-based
13. #giovanni_iacca @evostar_2012_málaga
Robot Base Disturbance Optimization
Trajectory plan: given a task, generate a trajectory (pos, vel, acc) for each joint so that
1) the end effector execute the task, 2) the trajectory is smooth, 3) kinematic/dynamic
constraint are satisfied, 4) the trajectory is optimal according to some criteria
14. #giovanni_iacca @evostar_2012_málaga
Robot Base Disturbance Optimization
Point-to-point problem:
define inter-knot points and
interpolate
(linear interpolation, spline, etc.)
Motion control: define the
torques to be applied
15. #giovanni_iacca @evostar_2012_málaga
Robot Base Disturbance Optimization
Free-floating
environment
Mutual disturbance between base and end-effector:
FB
= N-1
FE
↔ FE
= NFB
Fitness function: minimize the integral over
time of the norm of the acceleration vector
on the base
16. #giovanni_iacca @evostar_2012_málaga
Simulation Results
Matlab/Simulink implementation
{pos, vel, acc} x 2 knots x 3 joints = 18 variables
5th
order spline to model q(t), continuity condition on {pos, vel, acc}
4 different memory-saving optimization algorithms
Parameter setting suggested in original papers
30 runs x 10000 fitness evaluations
Wilcoxon Rank-Sum Test (confidence level 0.95)
17. #giovanni_iacca @evostar_2012_málaga
Simulation Results
Compact Differential Evolution (cDE)
Intelligent Single Particle Optimization (ISPO)
Non-Uniform Simulated Annealing (nuSA)
Without optimization (beginning of learning period)
With optimization (end of learning period)
18. #giovanni_iacca @evostar_2012_málaga
Conclusions
Some industrial applications are plagued by limited hardware
Compact algorithms, e.g. cDE, due to their compactness and
robustness are well suited for this kind of applications
We proposed a more efficient cDE (cDElight)
cDE proven to be successful on a complex space robotic
application
Possible alternatives and future works: different compact
frameworks (e.g. cBFO, cPSO), memory-saving MC