Progress for week 12 (2020)

Fra Robin

(Forskjeller mellom versjoner)
Gå til: navigasjon, søk
(Wonho)
(Budget)
 
(Én mellomrevisjon ikke vist.)
Linje 10: Linje 10:
** Turned out that last position bonus to fitness at the end of the run is enough to push the search in right direction
** Turned out that last position bonus to fitness at the end of the run is enough to push the search in right direction
** running experiments without it shows clear performance change for fitness only but similiar for ns only
** running experiments without it shows clear performance change for fitness only but similiar for ns only
 +
 +
[[fil:Comparison_of_Mean-Best_Distance_Traveled.png]]
 +
 +
** This suggests that MBF measurement is somewhat flawed, for beh.descriptor experiment - it has other consequence
 +
[[fil:Mean-Best_Fitness_Comparison_between_Behavior_Descriptors.png]]
 +
[[fil:Comparison_of_Mean-Best_Distance_Traveled_for_Behavior_Descriptors.png]]
* Static controller evolved with MOEA performing too well.  
* Static controller evolved with MOEA performing too well.  
** Considering only best genomes, it can handle unseen leg configurations and also walks as far as flexible ones(statistically insignificant).
** Considering only best genomes, it can handle unseen leg configurations and also walks as far as flexible ones(statistically insignificant).
-
** although fitness only run does perform better - in terms of walking gait & number of genomes managing all unseen leg configs - hard to directly compare as they are evolved in different EA
+
** although fitness only run does perform better - in terms of walking gait & number of genomes managing all unseen leg configs - hard to directly compare as they are evolved in different EA scheme
 +
 
 +
* While writing implementation, found out some flaws in controller implementation.
 +
** CTRNN output is limited to [0-1] due to sigmoid activation at output layer
 +
** At some point in the past it was scaled to [-pi/2, pi/2], which is input range DyRET env can take
 +
** this scaling has been gone missing whilst changing packages/merging/branching/testing
 +
** result is that leg movement is even more constrained than I have intended to. Possibly reason why static controller is also performing well on unseen leg configurations.
=== Budget ===
=== Budget ===
 +
* Continue writing

Nåværende revisjon fra 19. mar 2020 kl. 09:42

Innhold

Nikolai

Wonho

Accounting

Ran most of the experiments BUT

  • while comparing fitness only vs ns only vs moea - fitness only actually had highest fitness/distance traveled despite statistically not significant
    • Turned out that last position bonus to fitness at the end of the run is enough to push the search in right direction
    • running experiments without it shows clear performance change for fitness only but similiar for ns only

fil:Comparison_of_Mean-Best_Distance_Traveled.png

    • This suggests that MBF measurement is somewhat flawed, for beh.descriptor experiment - it has other consequence

fil:Mean-Best_Fitness_Comparison_between_Behavior_Descriptors.png fil:Comparison_of_Mean-Best_Distance_Traveled_for_Behavior_Descriptors.png

  • Static controller evolved with MOEA performing too well.
    • Considering only best genomes, it can handle unseen leg configurations and also walks as far as flexible ones(statistically insignificant).
    • although fitness only run does perform better - in terms of walking gait & number of genomes managing all unseen leg configs - hard to directly compare as they are evolved in different EA scheme
  • While writing implementation, found out some flaws in controller implementation.
    • CTRNN output is limited to [0-1] due to sigmoid activation at output layer
    • At some point in the past it was scaled to [-pi/2, pi/2], which is input range DyRET env can take
    • this scaling has been gone missing whilst changing packages/merging/branching/testing
    • result is that leg movement is even more constrained than I have intended to. Possibly reason why static controller is also performing well on unseen leg configurations.

Budget

  • Continue writing
Personlige verktøy