Skip to content

Latest commit

 

History

History
193 lines (131 loc) · 5.58 KB

TODO.md

File metadata and controls

193 lines (131 loc) · 5.58 KB
Raw

Items

tasks

the actually interesting bit, grow package by stepping through the tasks

  • x jumping goal: x by counting, x by average error
  • learn an e2p model, e.g. taxis with distinct goal- and proprio-spaces
    • clean up actinf/hebbsom and gmm's, ready to load with fit and predict
    • integrate these
  • actinf, imol, im proprio?
  • use kinesis as explorer and learn a model
  • artificial organism with brain equation: bacterial random search, directional field, internal simulation, TSPwAC, reservoir fillup/consumption
  • motor babbling
  • goal babbling

high-level design qu's

  • default configuration generator, also use it to repair/update old configs
  • revert to blocks paradigm, all blocks of same basic type
    • globally share config among all blocks
    • use one shared type "Block", init from config all the same
    • use one shared smdict to globally ferry information
    • every block lists the subscriptions and publishers it wants/creates
    • the config dict is mapped onto comp graph without structural change (smpblocks)?
    • hierarchical: "goal" is just another prediction in the motor command sense, such as proprio layer makes a prediction at proprio level, the lowest one, next level makes a prediction at one level up which becomes the goal for proprio, etc
  • test multiple robots, tasks, …
  • x how to separate world, robot, task appropriately: current state (v2) seems ok

add brains

  • interest models, learning from kinesis / kinesis as explorer
  • im, imol, actinf, eh, iso, …
  • hyperopt'ing
  • x taxis
  • x kinesis

add robots

  • logfile reading robot
  • sphero
  • cartpole
  • x stdr
  • x arm
  • x check pointmass kinematic vs. dynamic control: doesn't even make sense for kinematic control?

add worlds

  • add force fields (pm, sa)
  • add structure and obstacles (pm, sa)
  • add exteroception (pm)
  • x increase dimensions 3, 10

sm space

  • dimensions again: need to be able to access sm variables by name from anywhere (!!!)
  • partition order: make sure that the order of sm space partitions is consistent with respect to smdict keys etc
  • provide for an efficient way of integrating very high dimensional observations, like images without generating high dimensional column name arrays etc

dimensions

what is system, what is robot, what are dimensions, do proper spec of

  • proprioceptive
  • exteroceptive
  • interoceptive
  • reward system: pain, pleasure, hunger, …

fix plotting

  • distinguish analyses and plots
  • x plotting proper from prototype
    • x switch to pandas dataframes as structured arrays seem weird
    • x fix hardcoded tablename
  • x dataframes also solve the column name problem
  • x make sure plot funcs are dealing with figures and axes, so there can be a final aggregation of different parts into a final figure that can also be saved to print format

failsafe

  • make failsafe checks for configuration consistency

integration / robots lib

  • ideally, i would like to have a library of systems which can be equipped with wrappers for: explauto, ROS, … probably using decorators
  • x use explauto or not? no

experiment specification

  • how to define experiment structure in conf that can capture different example scenarios like:

    • single episode learning
    • multi episode learning (value func prop)
    • multi episode optimization (hpo, cma, evo, …)
    • infinite episode, …
    • single episode, single model, single task
    • single run multiple models single task
    • optimization run single model single task

  • IDEA: use a generic type "loop" which has a "step" method and a "stack" member, stacks being ordered dicts/lists of "loops"

  • hierarchy

    • 0th order innate hardwared controller (Darwinian) changes motor output according to rule
    • 1st order learning model (Skinnerian) changes controller parameters in order to change the change in motor output
    • 2nd order learning wraps around that again and changes the 1st order learners parameter to better change the 0th order parameters to change the motor output
    • it should be possible to apply the same learning principles on all those levels (kinesis, taxis, online model learner with exploration, actinf, evo, hyperopt)?

logging

  • how logging and ros publishing is the same or not: ROS also involves inputs whereas logging is only one-way
  • logging/publishing decorators
  • profiling log function, compare log2 (direct hdf5) + log3 (via pandas)
  • x column names for tables

efus

  • ultrastability
  • explorer (kinesis, interest models)
  • learner

dynamic challenges

  • dynamic creation of new variables
  • dynamic length of run
  • dynamic structural changes
  • make block types which get their step function body from the configuration

neural networks

  • include tricks of the trade foo: input cleaning, square augmented …
  • include deep learning foo

testing

  • do unit testing

submodules

which parts to do as submodules:

  • noiselib: uniform, gaussian, binomial, pareto, …
  • learnlib: rls, force, eh, iso, …
  • losslib: mse, mae, pi, ais, …
  • analylib: plot timeseries, plot histograms, plot hexbin, plot dimstack, scattermatrix, …

Filesystem

Separate the core component lib from the actual experiments, specification, and logging data.

Put that into smpexperiments.