Skip to content

Latest commit

 

History

History
178 lines (139 loc) · 6.94 KB

HISTORY.md

File metadata and controls

178 lines (139 loc) · 6.94 KB

History

Framework Improvements

February 10th, 2025 | By Thomas Breimer | PR approved by Thomas Breimer

  • Can save simulations as videos
  • Can decide to run in headless, video, window modes from terminal
  • Can run individuals
  • Can save individuals to videos
  • Added some plots

Feb 11th | By Abhay Kashyap

  • Added script to transform CMA-ES output to SNN weights and biases
  • Script reshapes the flat array into an array of shape (NUM_SNN * PARAMS_PER_SNN) and then returns a list of dictionaries explicitly separating weights and biases of each SNN

Introducing snn to evogym pipeline

Thurs. Feb 13th | Luodi Wang | SNN output to evogym through .json file

  • Saving the output into a json file from a mock (random) input
  • Deleted unused files: node.py, layers.py, net.py

Command Line Args Improvements

February 13th, 2025 | By James Gaskell & Thomas Breimer

  • Added proper named command line arguments for all scripts
  • Improved implementation with argparse

John's SNN Code

February 13th, 2025 | By Atharv Tekurkar

  • Removed unecessay elements
  • Fixed the functions that are giving errors
  • Added unit tests for the functions

Refactoring and Fixes

Feburary 15, 2025 | By Luodi Wang (partnered with Miguel Garduno on pylint)

  • get-output-state() method in the snn class from a py file
  • Within the SNN folder increased pylint score from 5.54/10 to 10/10 by 80.5%!

Feb 16th | Abhay Kashyap

  • Made the snn_parameters dictionary elements compliant with the way SNN's set_weights() method works
  • Completed testing unpacking tests with dummy cmaes output and setting them to SNNs

Refactoring

Feb 17th | Luodi Wang

  • Successfully completed robot configuration with its locations of corners from bestbot.json
  • Run inputs through the SNN
  • and we've now finally outputted actuator control values

Feb 17th | Abhay Kashyap

  • Merged all utility functions for SNN (eg. cmaes output => snn w&b; getting snn outputs, etc.) into one class
  • Updated SNN code to fix some duty cycle and firelog errors (duty cycle was always 0 as firelog was not getting updated because of threshold)

Feb 23rd | Abhay Kashyap

  • Removed files that were not being used and renames some files to more appropriate names
  • Refactored SNNRunner() into SNNController() and the class simpler to use for simulation - removed redundant functions and converted some functions into private functions for abstraction.

Feb 24th | Thomas Breimer

  • Completed integration

SNN Firing Experiments

Feb 25th | By James Gaskell & Matthew Meek

  • Added code to isolate a voxel and record the distances between robot corners
  • Output corner values for each time step to csv for analysis
  • Graphed expansion and contraction behaviors on a single voxel and large robot to determine SNN firing frequency

Feb 27th | Thomas Breimer

  • cmaes bug fixes

Mar 4th | Thomas Breimer

  • Added bestbot.json to snn-sim
  • Chnaged run_cmaes.py to use bestbot.json

Mar 4th | Guy Tallent

  • added run_cmaes2, run_sim2, morph2, actuator2. This maps the point massess to each voxel and allows you to get the distance from each of the active voxels to all four corners of the robot

Mar 4th | Jonathan Fischman

  • ring buffer implementation

Mar 5th | Abhay Kashyap

  • Created a log to track the frequency of target lengths obtained from SNN
  • Added notebook showing a simple visualization of this frequency
  • Added code to normalize inputs for better SNN computations

Mar 5th | James Gaskell

  • Created best_individual_latest.py which runs the best fitness from the most recent experiment or a specified gen number. Can be changed based on snn type
  • Implemented RSNN on new branch using CMA-ES taking into account n previous actuator values for each voxel

March 6th | Thomas Breimer

  • added run_experiment.py

Adding Ring Buffer

March 06th | Atharv Tekurkar

  • Added ring buffer
  • Implemented it for firelog in the snn structure
  • Removed duplicate snn files that were in cmaes_integration

Mar 8th | Abhay Kashyap

  • Added length method to ring buffer and made get() return n elements instead of all if parameter passed
  • Added logging of the membrane potential (level) to visualize the data accumulation and spikes

Mar 20th | Viet-Dung (Darren) Nguyen

  • Added implementation of latency tracking with 2 graphs of fitness over generation and velocity over simulation step

Ozempic TWO

April 3rd | Matthew Meek & Thomas Breimer

  • Removed folders with legacy code. voxel_length_test, rmhc_sim, cmaes_demo, morpho_demo.
  • Ozempic-ed cmaes_integration: removed all code non-essential for it to run.
  • Renamed cmaes_integration to cmaes_framework.
  • Reverted framework to old version circa week 9 Winter 2025

Windows

April 3rd | Thomas Breimer

  • Changed csv file naming for Windows compat

SNN Visualizations

  • Created feature branch to investigate SNN firing and levels
  • Forked repository and switched to SNN branch

April 6th | Hades Panlilio

  • Adjusted weights and neruon resets

Install instructions

April 7th | Matthew Meek

  • updated INSTALL.md for getting things working on the sudoless lab machines.

Pull Request Template

April 11th | John Rieffel

  • added pull request template

SNN Refactor

April 9th | Thomas Bremer

  • Added docs for all SNN methods
  • Fixed neuron duty cycle
  • Added logs for duty cycle and proper fire log
  • Added option to generate SNN logs with run_individual.py
  • Added graphing notebook for SNN logs

Apr 10th | By Takumi Kojima

  • bugfix of best_individual_latest regarding the directory path
  • change the default setting best_individual_latest

best_individual changes

April 14th | James Gaskell

  • Edited best individual to create logs
  • Updated requirements for the correct version of matplotlib
  • Fixed symlink for MacOS

Data file structure

April 16th | James Gaskell

  • Edited the data file structure to include symlink for latest data and grouping of videos, logs and genomes
  • Edited creation of logs - one log per genome for tidiness

Devops instructions

April 25th | Takumi

  • Added DevOps.md which is a manual of how to use screen command, du command, and ssh to lab computers.

Data file structure continued

April 25th | James Gaskell

  • Edited file structure to collect multiple runs under the same experiment name
  • Edited plotting to show fitness over time for mutliple runs of the same experiment
  • Moved around data and symlinks to keep everything in the data folderg

Multiple hidden SNN layers

April 28th | Hades Panlilio

  • Added parameter hidden_layers to specify amount of hidden layers/neurons per layer
  • Added hidden layer functionality to SpikyNet
  • Updated files that run the simulation to accept new parameter and multiple layers

run_experiment Update & parameterize spiky decay and robot config

April 28th | Takumi

  • Updated run_experiment with all possible params for now
  • parameterized spiky decay
  • parameterized robot_confi

Supporting alternative SNN Input

April 29th | Thomas Breimer

  • Added infrastructure for alternative SNN input options
  • Added distance to all other actuators SNN input option