Having tried it, I found the evolution method a little lacking. I've seen some happenchance good solutions leap ahead (literally and/or figuratively) in one generation, but because of the "recombine and start afresh with the new generation" whatever spark of genius there was gets lost in the next generation. Maybe the key bits were either side of the 'splice point', however it stores its 'code' for operation, and without one bit now working to support the other and vice-versa (or actively acting to sabotage the combined effort) you get offspring more dud than either of the parents.
If not too dud, then maybe the children survive to pass on the extremities of their parental Brains to the next generation, with recombination up the line of generations melding them both together again (with mid-Brain info being a mix of influences as other lines carry through via the jumping resplicing points) but it's a bit disappointing.
Many creatures IRL stick around to compete (beyond any parental duties their species commits to) with their offspring and produce further generations too, so their line isn't lost if their first kids turn out to be unfit (especially compared with themselves), at least as long as old age and bad luck hold off their icy hands of post-mortality.
For the Darwinian Poetry evolver, the scheme was that Gen0s (unrelated randomised constructs) started out the competition, with the most favoured 'poems' being bred amongst themselves for Gen1 offspring (which was also done with splicing, but it was multiple splices1) which were then added to the population. As the population grew, culling (removal from the active pool, still viewable) was enacted upon the consistently bad performers (who may or may not have descendents by this point) without any reference to age.
In fact, there was no compulsory senility/die-off. And no bar at all to (currently) successful entities from different generations breeding (GenX+GenY => two complimentary Gen{max(X, Y)+1} offspring, for recording purposes). If a Gen0 had turned out to be a masterpiece, already, it would have competed its way to sire offspring for as long as the 'environment' didn't just get bored of seeing it and voted it down just 'because'.
I wonder if I still have a readable archive of my population scrapes, anywhere? Probably not, but I'll check some old CD-Rs that have doubtless deteriorated and needn't take up space on my shelves any more.
Anyway, this was with a working population (at any given time) far in excess of ten. Hundreds. Maybe thousands. Could be beyond this Evolving Stick-Things' capacity, having to also simulate them all to test their capabilities. But culling the worst half then breeding up enough replacements (needs multiple-of-four constructs at a time? - or just adjust the 'half to cull' one way or another to make it an even number swapped out and created in their place) might have been my approach. (In fact, it has been, indifferent little toys I have made myself, for my own interest.)
Forgot to add: Gen0s are always going to be overwhelmingly failures (i.e. lucky if they don't fail badly, virtually unknown that a positive outcome arises from the initial RNG Creator), and ten generations (of a constant ten 'breed then die' populous, with no way of looking for sort-of-alike similarities between parents and just the one switchover splicing point, for all the reasons give above) just isn't going to be enough for anything complicated. The challenge for Running seems to be barely susceptible to serendipitous selection of a decent hobbling (minimalist) creature. I haven't checked jumping - that might be easier to get good at (just synchronise things to get a good 'launch'). The stairs look lethal.
And wouldn't it be evolutionarily interesting to also vary (by mutation, at first, then inheritance) the body geometry. Shift the joints around slightly, for differing bone lengths and initial inter-bone angles. Even adjust the muscles from equidistant stretch and shorten to a centre-offset range. Adds complexity, but adds potential. Like getting the length/range of limb your stair-climber deserves, rather than the one you designed it with, without having to restart after watching the failure to improve and moving the joints around yourself with very nearly as little expertise, then starting the evolving from scratch again. An option? Probably already suggested over on that site's forum, but prior to my going there I thought I'd unload here.
1 The same number per parent, but with independent placement on each parent, so the first two words of one parent might swap with the first four of the other, and so on, changing poem lengths. I think this was revised later, but it allowed Selection to favour longer or shorter poems, and (a couple of generations down the line, or more) repeated phrases to creep in. Obviously for a static-sized 'Brain' (one assumes it's a rigidly templated NN, with a fixed depth and a given width of possible connections between each layer betwixt sensory and motor node layers - doubtless explained somewhere on the forums for the thing, which I must look at later) you'd just generate the single set of multiple splice'n'switch points and keep the data the same size. But it gives the possibility of a good sensory cortex (one end of the layers) and a good motor cortex (the other end of the layers) to inherit from the other side of its parentage a novel mid-brain 'remixer', etc.
Author
Topic: Evolution, simulate a creature of muscle and bones (Read 3291 times)