Source

bubble-economy / TODO.rst

Full commit

TODO

Action plan

Sub-lists are are in descending priority order.

# Absolute top priority task killing finality

# work out why velocities all appear to be in the same quadrant for quite
high noise runs

# bootstrap everything # keep trial stat arrays around in summaries so that R can do the

aggregation

# generate the final ultimate kernel density estimates of behaviours in R # have different number of samples per-axis

# Faster simulation

# throw out some normalisations- just make sure that random vectors are on
average unit length.

# use gridding to bin cells into neighbourhoods

# wilder interpretations

# phase transition against the percent of distance each timestep goes to
randomness

# warp P1 axis into angular noise analog (Finding it analytically?)

# Faster, more powerful MI

# use a priori binning # worry about missing bins for, e.g. correctors # go to CUDA for mi calcs? - http://users.cecs.anu.edu.au/~ramtin/cuda.htm # use binless inference? (kernel density? spline? k-neighbour?) # does a kd tree give me MI for free? # use Cython to wrap NSB_entropy

properly instead of calling as a subprocess

# Better stats

# 2d graphs throughout for gridded sims # use bootstrap estimator # make sure those "Warning: Null output conditional ensemble for output

foo" don't reflect bad params

# evaluate using velocity deltas # test on other phase transition data # consider stationary vs non-stationary # check susc. stat. Doesn't seem to be the same as order std dev. # consider ergodic walks and independence

# Better sampler

# importance sampling # persist job management state across crashes in the picloud dispatcher.

# Better visualisation

# animation of acting agents

# as slices of a 2d/3d thing # as a 3d animation (see also)

# mark lines for B&W using plot

# experiments to do

# higher dimensional behaviour # time window sweep # small-number-of-agents analysis

Axis interpretation

  • valuations of equities
  • buyin' and sellin' agents
  • equities (must conserve)
  • total market value.
  • think about stock market in terms of axes and dimensions - equities are close if they have similar historical price movements - that is, how good they have been as predictors of each other in the past - rather than current price value. and price changes come from stock movements. How would one link equities and traders in this fashion?

Pie-in-the-sky

  • explore alternative distance measures that allow similarity in a small number of dimensions (Cosine distance? correlation distance? extended jaccard distance?)
  • scale number of agents to keep Pi_4 constant in higher dimensions
  • add option to remove self-adjacency from calculations at all stages and explore consequences
  • sanity-check minkowski distance
  • handle more boundary constraints * esp. bounded-below-only boundary conditions
  • calculate prices using demand curve
  • integrate Sumatra
  • integrate Ruffus
  • better test coverage
  • document for posterity
  • contribute back to pyentropy
  • export to a gephi animated graph