x:trade-only, 3x3 4a

[jarbus]

change average fitness over all batched rollouts to max fitness over batched rollouts

[jarbus]

and change network architecture for smaller grid

[jarbus]

add more gens

[jarbus]

lower mutation rate

[jarbus]

if this doesn't work, try 1x1 2a

[jarbus]

running on p2k-4a.arg

[jarbus]

seems to not work. next step is to run a 1x1 2a experiment

[jarbus]

maybe experiment with large quantities of rollout groups?

[jarbus]

for this, batch size of 1 can work. in the 1x1 case, I'm noticing a novelty collapse, likely due to the small num of elites which optimize for fitness intiially. perhaps maintaining more diversity will help this. I should also not decimate the exploration param, probably.

[jarbus]

maybe need to reduce the amount that agents trade? or make an agent's food source be deterministic?

[jarbus]

moving this thread to 1x1