> Monotonic-only improvement is the path to getting trapped. Take chances, make mistakes, and get messy.
The evolution disagrees.
The evolution disagrees.
All major trophic level breakthroughs are powered by evolving a reserve of efficiency in which mult-step searching can occur.
multicellular life, collaboration between species, mutualism, social behavior, communication, society, civilization, language and cognition are all breakthroughs that permitted new feature spaces of exploration that required non-locally-optimal transitions by the involved systems FIRST to enable them.
Trust is expensive and can only get bought in the presence of a surplus of utility vs requirements.
Basically, intelligent behavior is optimizing for "future asymptotic entropy" vs maximizing any immediate value. How intelligent a system is then become a measure of how far in the future it can model and optimize entropy effectively for.
(updated with pdf link)
[1]: Thermodynamic Game Theory: https://adamilab.msu.edu/wp-content/uploads/AdamiHintze2018....
[2]: piKL - KL-regularized RL: https://arxiv.org/abs/2112.07544
[3]: Soft-Actor Critic - Entropy-regularized RL: https://arxiv.org/abs/1801.01290
[4]: "Soft" (Boltzmann) Q-learning = Entropy-regularized policy gradients: https://arxiv.org/abs/1704.06440
But your comment was refreshing, could you briefly expand on the "multicellular" life part? Did you mean that it enabled more non-locally-optimal transitions, or that it required them to appear?
> Take chances, make mistakes, and get messy.
But then seemed to indicate evolution disagrees.
I might be misunderstanding your point, but it sure seems like, evolution tries a bunch of stuff, and whatever reproduces kinda wins.
That seems like, take chances, make mistakes, get messy. That seems like the core of evolution.
Could you clarify or refine what you’re saying? The two seem at odds.
I got inspired by this article: https://writings.stephenwolfram.com/2024/05/why-does-biologi...
I used to do an "optimization" on my genetic algorithms. I'd ensure the highest scoring genome of the last population was a member of the new one. It made sure every single generation improved or stood still.
It was a good idea to keep a copy of the "best" genome around for final output, but by keeping it in the search space, I was damaging the ability of the algorithm to do it's job by dragging the search space constantly back to the most recent local optima.
Takes a lot of luck to evolve cooperation multiple times at once, much more likely to happen in a situation where the selection pressure is lower, not higher.
The thing about evolution is that you are sampling many times in different directions. So "luck" isn't that hard to achieve.
> Pretty sure those non-cooperative strategies quickly burn themselves to extinction though.
Um, most life hanging out in the same tropic level or lower is heavily predated upon. Competition is the norm.
Luck is hard for cooperation because it is a coordination problem. You basically have to evolve cooperation entirely as a unexpressed trait then trigger it in the population almost simultaneously. The mechanisms of cell cooperation are critical dividers on our evolutionary trees for a reason, they are rare and dramatic in consequence. Cell populations regressing in terms of coordination behavior (see cancer) is one of their most problematic failure modes and it is only very weakly selected against.
I'm referring to the predator-prey population cycles. If you overexploit your prey you are going to run out of food and see your population thin out rapidly from starvation. Hence hyper-competitive strategies would get outbreeded by less competitive but sustainable strategies.
High predation levels would require equally high cooperation levels amongst prey to ensure rapid reproduction to sustain the food supply. If we go down the food chain it's the same thing, plant life, celluar life, etc, has to be flourishing to sustain the upper levels.
My point was there's no pressure without constraints. A faster-reproducing species will only apply a pressure if starts exhausting a resource or similar.
> The evolution disagrees.
It’s not an either/or. Vast modularized localized improvement allows for the ability to prune and select what does and doesn’t work.
Ah yes monotonic-only improvement by way of making every small, messy mistake possible and still probably going extinct is definitely the way to go