jarbus / evo Goto Github PK

i suspect this could be fixed by always including the best fitness as a copy in the pop

edit: occurs in gesmr-adapt={noelite1,elite1}

testing if a GA can learn to forage. may need to do an hparam sweep. This should
be doable.

this should optimize over different fitnesses and novelties

this probably doesn't matter except for minor performance details

trace paths of agents
thickness of path corresponds to something
circle pick/place actions

arXiv:2204.04817

evenly-spaced triplets have significant computational overlap

(using large model, 2.4M params)

model creation vs run batch

julia> @btime models = Dict("f0a0"=>re(reconstruct(nt, UInt32.([3, 4, 5]))))
3.576 ms (436 allocations: 28.27 MiB)
julia> @btime rewards, mets, bc = run_batch(env, models, args)
472.274 ms (126467 allocations: 335.63 MiB)

model creation scales with seed count

julia> @btime models = Dict("f0a0"=>re(reconstruct(nt, UInt32.([rand([1,2,3]) for _ in 1:100]))))
  60.308 ms (536 allocations: 28.28 MiB)

re adds negligible overhead

julia> @btime reconstruct(nt, UInt32.([1,2,3,4,5,6,7,8,9,10]));
  6.779 ms (7 allocations: 18.83 MiB)
@btime re(reconstruct(nt, UInt32.([1,2,3,4,5,6,7,8,9,10])));
  7.453 ms (431 allocations: 28.26 MiB)

launching \sigma grid search = {0.0001,0.0005,0.001,0.005} at 1:18am

100 processors each

todo: average times for each processing step

run experiments with/without:

adaptive mutation rate
adaptive exploration rate
mut-mod from world model neuroevolution paper
run experiments to confirm that these work

run an exp with single-agent trade

also add mutation rate mean, sd, min and max to metrics

paper details:

takes 200 generations for a pop size of 20k to solve hard maze, 6 hours
on their

by generation 5, there is significant policy diversity around the origin

el 60

added assertion to make sure elite stays in next pop, if it still collapses then idk