Commits

Commit history

1. Commits on Jun 25, 2018

   • Updated ML/Agents to v0.4

     MarcoMeter

     committed
     a2c54b7View commit details

     Copy full SHA for a2c54b7

     View code at this pointBrowse repository at this point
2. Commits on Apr 27, 2018

   • (WIP) ++

     MarcoMeter

     committed
     6553382View commit details

     Copy full SHA for 6553382

     View code at this pointBrowse repository at this point
3. Commits on Apr 25, 2018

   • (WIP) SCC uses visual observation for now and discrete actions

     MarcoMeter

     committed
     b455d54View commit details

     Copy full SHA for b455d54

     View code at this pointBrowse repository at this point
4. Commits on Mar 26, 2018

   • BallLabyrinth: Fixed collision/distance bugs, improved inputs (relative position instead of normalized direction, removed distance, fixed normalization of velocity), adjusted hyperparameters

     MarcoMeter

     committed
     7f4b66eView commit details

     Copy full SHA for 7f4b66e

     View code at this pointBrowse repository at this point
5. Commits on Mar 25, 2018

   • Intermediate progress on BallLabyrinth (which is still to be tuned due to the ml-agents update)

     MarcoMeter

     committed
     5c277d9View commit details

     Copy full SHA for 5c277d9

     View code at this pointBrowse repository at this point

   • Ignoring unity-environment.log

     MarcoMeter

     committed
     d5a2edcView commit details

     Copy full SHA for d5a2edc

     View code at this pointBrowse repository at this point
6. Commits on Mar 23, 2018

   • SCC: Reduced inputs to 4 and improved hyperparameters.

     MarcoMeter

     committed
     a3b3953View commit details

     Copy full SHA for a3b3953

     View code at this pointBrowse repository at this point

   • Major update on simple continuous control:

     Show description for 600e0f4
     MarcoMeter

     committed
     600e0f4View commit details

     Copy full SHA for 600e0f4

     View code at this pointBrowse repository at this point
7. Commits on Mar 22, 2018

   • DC2D: All actions are discrete now. Two more actions are introduced which combine rotating and shooting. Shooting is penalized by a negative reward. The observation of the adgent still needs some a…

     Show description for 8ca17c4
     MarcoMeter

     committed
     8ca17c4View commit details

     Copy full SHA for 8ca17c4

     View code at this pointBrowse repository at this point

   • Updated DestroyComets2D to ml-agents version 0.3b

     MarcoMeter

     committed
     1cb1833View commit details

     Copy full SHA for 1cb1833

     View code at this pointBrowse repository at this point

   • Updated simple continuous control to v0.3b

     MarcoMeter

     committed
     e4f5fe8View commit details

     Copy full SHA for e4f5fe8

     View code at this pointBrowse repository at this point

   • Updated python folder to ml-agents 0.3b

     MarcoMeter

     committed
     106c046View commit details

     Copy full SHA for 106c046

     View code at this pointBrowse repository at this point

   • Updated BallLabyrinth to ml-agents version 0.3b (curriculum lessons don't kick in right now)

     MarcoMeter

     committed
     287e07aView commit details

     Copy full SHA for 287e07a

     View code at this pointBrowse repository at this point
8. Commits on Mar 19, 2018

   • Updated python folder to 0.3v and BasketCatch environment.

     MarcoMeter

     committed
     9286697View commit details

     Copy full SHA for 9286697

     View code at this pointBrowse repository at this point
9. Commits on Mar 9, 2018

   • Updated ML-Agents to 0.2.1d, Unity update to 2017.3.1f

     MarcoMeter

     committed
     170a058View commit details

     Copy full SHA for 170a058

     View code at this pointBrowse repository at this point
10. Commits on Jan 15, 2018

    • Ball Labyrinth: Ignoring curriculum learning for, each agent has its own start position which is based on the available curriculum lessons (training results are still pending)

      MarcoMeter

      committed
      770cb3fView commit details

      Copy full SHA for 770cb3f

      View code at this pointBrowse repository at this point

    • Modified the layer structure for using image input

      MarcoMeter

      committed
      1afbe81View commit details

      Copy full SHA for 1afbe81

      View code at this pointBrowse repository at this point

    • DC2D: removed shooting cooldown, added functionality for using discrete actions, expanded the state space by the velocity of detected comets (no proper training results yet, rather spray and pray b…

      Show description for 0072b7f
      MarcoMeter

      committed
      0072b7fView commit details

      Copy full SHA for 0072b7f

      View code at this pointBrowse repository at this point

    • Possibility to plot observation in SSC PPO

      MarcoMeter

      committed
      b42d147View commit details

      Copy full SHA for b42d147

      View code at this pointBrowse repository at this point
11. Commits on Jan 8, 2018

    • Ball Labrinth: Disabled camera observation for now and created 43 continuous inputs instead. Training reaches lesson 5 of the curriculum now.

      MarcoMeter

      committed
      3a7b729View commit details

      Copy full SHA for 3a7b729

      View code at this pointBrowse repository at this point
12. Commits on Dec 25, 2017

    • Simple Continuous Control: Results are now reliable for a 64x64 greyscale image input (converges at ~100k steps)

      MarcoMeter

      committed
      c7f46d2View commit details

      Copy full SHA for c7f46d2

      View code at this pointBrowse repository at this point
13. Commits on Dec 23, 2017

    • Ball Labyrinth: Agent is now able to get to lesson 4 after 600k steps

      Show description for 66094b1
      MarcoMeter

      committed
      66094b1View commit details

      Copy full SHA for 66094b1

      View code at this pointBrowse repository at this point
14. Commits on Dec 18, 2017

    • Updated documentation to add info about the last introduced environments

      Show description for 82d563f
      MarcoMeter

      committed
      82d563fView commit details

      Copy full SHA for 82d563f

      View code at this pointBrowse repository at this point

    • Completed the flow of the training process based on curriculum learning (I a model is not trained yet)

      MarcoMeter

      committed
      806b611View commit details

      Copy full SHA for 806b611

      View code at this pointBrowse repository at this point
15. Commits on Dec 11, 2017

    • - Updated ML Agents to 0.2

      Show description for 2c42ec0
      MarcoMeter

      committed
      2c42ec0View commit details

      Copy full SHA for 2c42ec0

      View code at this pointBrowse repository at this point
16. Commits on Dec 4, 2017

    • Each environment is now a seperate Unity project.

      Show description for 16d88eb
      MarcoMeter

      committed
      16d88ebView commit details

      Copy full SHA for 16d88eb

      View code at this pointBrowse repository at this point
17. Commits on Dec 3, 2017

    • Minor adjustments, which led to the first more reasonable behavior

      Show description for 9632df0
      MarcoMeter

      committed
      9632df0View commit details

      Copy full SHA for 9632df0

      View code at this pointBrowse repository at this point

    • Added "Destroy Comets 2D" envinronment (training doesn't work out yet)

      MarcoMeter

      committed
      69de8f4View commit details

      Copy full SHA for 69de8f4

      View code at this pointBrowse repository at this point
18. Commits on Nov 15, 2017

    • removed Unity's README.md on using python

      MarcoMeter

      committed
      74f1024View commit details

      Copy full SHA for 74f1024

      View code at this pointBrowse repository at this point

    • Initial commit of the environment "Basket Catch" (BC)

      MarcoMeter

      committed
      2e8f7a0View commit details

      Copy full SHA for 2e8f7a0

      View code at this pointBrowse repository at this point
19. Loading