Progress for week 6 (2020)

Fra Robin

(Forskjeller mellom versjoner)
Gå til: navigasjon, søk
(Wonho)
 
(Én mellomrevisjon ikke vist.)
Linje 5: Linje 5:
== Wonho ==
== Wonho ==
=== Accounting ===
=== Accounting ===
 +
* new behaviour descriptor implemented - last position
 +
** the results are more promissing - 4 evolution processes generated same walking gait, low crawling ( this needs modification on fitness function )
 +
** Needs longer evaluation for each genome to be effective
 +
* Vanila NEAT - generates very uncoordinated and unstable gaits
 +
* Tested some closed loop control scheme - setting its foot pattern or current joint angle as input to the network ( Risi uses current joint angle of upper joints )
 +
** The results are mostly devastating - systems converging to stable state ( not moving at all ) after few iterations.
 +
** This used to work okay with MultiNEAT's CTRNN - perhaps there are some difference in implementation in neat-python
 +
* Implementing MOEA
 +
** Tested Pareto front selection scheme on novelty archive ( currently limiting number of individuals since computation time increases exponentially )
 +
** The results doesn't seem very different from before
 +
** Will implement this for NEAT reproduction process
 +
* All experiments are recording git sha in the logger so that it can be regenerated later on
 +
=== Budget ===
=== Budget ===
 +
* Finish implementing MOEA in population reproduction process
 +
* Create some graphs from recent experiments
== Tony  ==
== Tony  ==
Linje 13: Linje 28:
=== Accounting ===
=== Accounting ===
Hvordan kan dette på best muligvis overføres til vinkel/vinkelhastighet?  
Hvordan kan dette på best muligvis overføres til vinkel/vinkelhastighet?  
-
theta = w*f(x) + w*g(x) + ... hvor da f(x) og g(x) er motion-features kurver
+
theta = w*f(x) + w*g(x) + ... hvor da f(x) og g(x) er motion-features kurver pluss mange i tillegg
[[Fil:Motion_features.png]]
[[Fil:Motion_features.png]]

Nåværende revisjon fra 6. feb 2020 kl. 09:36

Innhold

Nikolai

This week's news

  • Har fått hjelp av Ulysse og kan nå hente ut EMG data i Unity.

Wonho

Accounting

  • new behaviour descriptor implemented - last position
    • the results are more promissing - 4 evolution processes generated same walking gait, low crawling ( this needs modification on fitness function )
    • Needs longer evaluation for each genome to be effective
  • Vanila NEAT - generates very uncoordinated and unstable gaits
  • Tested some closed loop control scheme - setting its foot pattern or current joint angle as input to the network ( Risi uses current joint angle of upper joints )
    • The results are mostly devastating - systems converging to stable state ( not moving at all ) after few iterations.
    • This used to work okay with MultiNEAT's CTRNN - perhaps there are some difference in implementation in neat-python
  • Implementing MOEA
    • Tested Pareto front selection scheme on novelty archive ( currently limiting number of individuals since computation time increases exponentially )
    • The results doesn't seem very different from before
    • Will implement this for NEAT reproduction process
  • All experiments are recording git sha in the logger so that it can be regenerated later on

Budget

  • Finish implementing MOEA in population reproduction process
  • Create some graphs from recent experiments

Tony

Budget

Prøve på mobile robot med 2 ledd. (ikke 4 DOF)

Accounting

Hvordan kan dette på best muligvis overføres til vinkel/vinkelhastighet? theta = w*f(x) + w*g(x) + ... hvor da f(x) og g(x) er motion-features kurver pluss mange i tillegg

Fil:Motion_features.png

Personlige verktøy