Content pfp
Content
@
0 reply
0 recast
0 reaction
Breck Yunits pfp
Breck Yunits
@breck
ETA!: A Measure of Evolution https://breckyunits.com/eta.html
1 reply
0 recast
0 reaction
Connor McCormick ☀️ pfp
Connor McCormick ☀️
@nor
Why would reducing the size of the assembly pool increase the lifespan of an idea?
1 reply
0 recast
1 reaction
Breck Yunits pfp
Breck Yunits
@breck
Increase the lifespan of a *bad* idea. A larger assembly pool means a bad idea will evolve into a good idea faster.
1 reply
0 recast
1 reaction
Connor McCormick ☀️ pfp
Connor McCormick ☀️
@nor
why would that be the case? A larger assembly pool just means the search space is larger but it says nothing about the average fitness or distribution of possible assemblies
1 reply
0 recast
1 reaction
Breck Yunits pfp
Breck Yunits
@breck
If A' is an alteration of A with a quantum jump in fitness, and A' is composed of (A + Z), then we can't get to A' until Z is in the assembly pool.
1 reply
0 recast
1 reaction
Connor McCormick ☀️ pfp
Connor McCormick ☀️
@nor
Ok but if I add X & Y then the assembly pool is larger but won’t reach A' because A' requires Z I can do this infinitely, adding elements to A that don’t improve fitness Larger A increases the search space with no guarantee of improved fitness, in fact increasing the lifespan of “bad” ideas, not decreasing
1 reply
0 recast
1 reaction
Breck Yunits pfp
Breck Yunits
@breck
There is a guarantee of improved fitness by increasing A. If you only added bad ideas to A, then that implies you have a filter F that can detect good ideas, so just add !F. If you add all ideas at random, then you will add ideas that improve fitness.
1 reply
0 recast
1 reaction
Connor McCormick ☀️ pfp
Connor McCormick ☀️
@nor
Haha. All you're saying is that there's a guarantee of improved fitness by increasing A as long as subassemblies on average improve fitness. That's a tautology. What this formalism is missing is that the context / environment informs the time-to-fitness. Here, you can test it easily: implement a genetic algorithm searching for the string "abababababababa". Now, randomly select a subset of letters from the English alphabet as the assembly elements A (e.g. A = {c, g, e, z}). On average, time-to-fitness will tend to decrease with each marginal element added to the assembly. I'd guess the graph looks something like this:
1 reply
0 recast
1 reaction
Breck Yunits pfp
Breck Yunits
@breck
oh you just gave me a really big idea. brb.
1 reply
0 recast
1 reaction
Connor McCormick ☀️ pfp
Connor McCormick ☀️
@nor
P.S. said this backward "On average, time-to-fitness will tend to decrease with each marginal element added to the assembly." but the graph is correct. I.e. there comes a point where starting a genetic algorithm with a larger number of elements on average slows down its "fitness rate" / learning rate.
1 reply
0 recast
1 reaction
Breck Yunits pfp
Breck Yunits
@breck
The bigger idea you gave me is unrelated and may take me a few days. As to your point, the chart is not correct, as you are assuming serial testing, but there is no slowdown from adding to the assembly pool, because every combination is tested simultaneously.
0 reply
0 recast
0 reaction