Slither.io-bot icon indicating copy to clipboard operation
Slither.io-bot copied to clipboard

Published 20 hours ago • verified publisher icon ermiyaeskandary

Evolution rating for each version

Open Evanito opened this issue 8 years ago • 5 comments

Hey guys, so I've been thinking.

When it comes right down to it, what are we trying to make here? The answer: a bot. And what do we really want that bot to do? Answer: be as long as possible.

We want the bot to be able to get as long as possible.

How will we know if we have made progress? My idea: create a second script (named evolutiontest.js) that does a few simple things:

  • Once initiated manually, start the bot with uniform settings (Settings agreed to be standard.) and disable player controls.
  • Run the bot like normal, until death, where it records the maximum length (and possibly survival time) to a list.
  • Repeat several times to get an accurate feel of a real survival situation. (My statistics experience tells me we will need at least 30 trials to be able to use the central limit theorem and outweigh outliers.)
  • After the trials are completed, record the data to the github comment being tested. (Include average length, avg lifetime, avg FPS , browser). This step will be posted to Github manually.

With this process, I feel like we can verify if a commit had a negative effect on the performance of the bot where it really matters: its efficiency.

Evanito avatar May 11 '16 01:05 Evanito

maybe the second version slimmed ?

Seple avatar May 11 '16 01:05 Seple

Sounds like we really need this done soon so we can add it to the PR approvals workflow.

ermiyaeskandary avatar May 13 '16 14:05 ermiyaeskandary

Honestly, getting this in opens a huge avenue for things like genetic algorithms, which, in my opinion, are amazing.

dracco1993 avatar May 15 '16 02:05 dracco1993

@dracco1993 I would think a machine learning algorithm would be more advisable. Given the design of a genetic algorithm the goal of the snake is to become larger of which is not a predefined method. It requires a sense of intelligence. By using a machine learning algorithm it can slowly learn how close is too close, how far from the center is good, maybe what constitutes a good goal for food and many other variables that can be questionably assigned. Another does side to a genetic algorithm is I would say performance. If the genetic algorithm is used for route planning, as that being the only implementation in which I see it able to be applied to, it would have to re-evolve after every state change.

ghost avatar May 22 '16 03:05 ghost

http://synaptic.juancazala.com/ http://cs.stanford.edu/people/karpathy/convnetjs/ javascript neural network libraries

deepbluesea avatar May 30 '16 09:05 deepbluesea