Skip to content

Latest commit

 

History

History
72 lines (54 loc) · 2.05 KB

README.md

File metadata and controls

72 lines (54 loc) · 2.05 KB
Raw

Adabot - Evolving ANNs and FSMs

Build with:

cd src/
./build build ugv

Run with

cd src/
./build run ugv

The build script adds $HOME/.local/include to the include path, and $HOME/.local/lib to the DYLD_LIBRARY_PATH variable.

Revisit Logger

Use the following to add the logger submodule:

git submodule add revisit/logger
git submodule update --init
cd logger
git remote rm origin
git remote add origin https://github.com/revisit/logger.git
git push origin master

To update the submodule, use:

cd revisit-logger
git pull

Generate Visualization/Trajectory/Evaluation Data

  1. Pick individual from evolutionary history
    • high negative fitness from bottom of a outcmaesxrecentbest.dat file
  2. Copy genome (in range 0 to 10)
    • all values from fitness to end of line
  3. Run experiment with --genome flag
    • ./evolve_ugv_bnn.py --genome "2.68489637316 9.99723776033 2.83092512356 0.743949436278 1.45031247252 9.84016036964 1.02868840043 4.73290186541 1.72245190821 9.72243287632 2.21900453463 6.49723985829 9.34603023206 9.29940244589 3.43749814329 7.86080041933 0.486365856971"
  4. (Compare fitness with expected)
  5. Copy output string (real values)
    • these are genome values converted to simulation values
  6. Run *{_vis|_eval} binary
    • ../bin/ugv_bnn_eval '30 0 0 0.1015 0.16 0.02283 0 0 3.872 -3.177 -0.2137 -2.622 3.778 -2.225 1.198 3.477 3.44 -1.25 2.289 -3.611' > tmp.csv
    • first value is TIME_STOP
    • second value is NUM_OBSTACLES
    • third value is OBSTACLE_SEED
  7. Output to animation file

BNN Sweep

  1. Pick individual from evolutionary history
  2. Copy genome (in range 0 to 10)
  3. Run experiment with --genome flag
  4. (Compare fitness with expected)
  5. Copy only BNN parameters from output string (real values)

TODO

  • template NN and BNN

  • really need to convert UGV to a class (tons of code duplication)

  • refactor controllers into separate files

  • Get good trajectory/visualization controllers from evaluate runs (note the seed even though the spacing will be different).