Some Ideas

[blattsand]

I have been working on Programms similar to this one.

I am not sure if these Ideas are implemented in any way.

CTRNN Continuous Time Recurrent Neural Networks, seem to be the best fit for these kind of Simulations, single dual or even tripple layered.
Also Random Numbers as Input, is an Efficient way for them to find a random path to follow.
Use of Senses on or off, they should be able to turn off senses for energy-use-reduction, this way blind bots with hearing could form groups and show a specific way of life.
I also think that spike costs should be added.
Input into the neural network from the output of the same can be very helpfull for the bots, if this is not yet done.
Waste products, i believe Waste products should be implemented, I have seen it far too often groups of sitters, bots that just sit around and wait for food to come to them, this should be ineffektive but it isnt in highly populated herbivore worlds.
I believe Trees should be used for saving of the bots, for more efficient distance calculations, it should be possible to process more bots this way, I was able to process about a world of 9000 bots with CTRNN Brains ranging to 200 Neurons with acceptable speed, so there is more potential.

However I am not sure if any of these Ideas is implemented, I find it so hard to go through other coders code :-D.

Sorry for missspellings. Am only part English.

Hope you Enjoy
Me a.k.a. Noodles

[Andrej]

Hey, thanks for comments I agree with a lot of this!

- Not familiar with CTRNNs but I tried to implement the intuition of something similar. Every neuron has a target value that it computes, but the ACTUAL output only slowly converges towards the target value over time. So it's a tiny dynamical system. Should look into this more.

- There is a cost for spike and movement and so on, but not for using senses right now. That's interesting idea to actually add this for all sensors as well.

- The problem with sitters is a big one and I hate it. Basically it's bots who hang out at one spot and feed off things that happen to die near them. I think this is primarily a problem because I can't simulate a big world. Which brings me to:

- Optimizing the distance calculations and scaling up the size of world. Definitely using kd-trees etc for all of this and so on. Absolutely agree and this would be what I would implement next.

Thanks for the comments! I wish I had time to work on this :D Maybe I can sneak in a few hours this weekend and clean up the mess the code is right now, and begin adding some optimizations for scaling this all up.

Andrej

--
You received this message because you are subscribed to the Google Groups "Scriptbots" group.
To unsubscribe from this group and stop receiving emails from it, send an email to scriptbots+...@googlegroups.com.
For more options, visit https://groups.google.com/groups/opt_out.

[blattsand]

I have read myself into this alotin the past days since.

I believe k-d-trees is not a solution, since it has to be balanced to be effektive, and deletion and insertion destroys balancedness. But and this is a big BUT you simply have to readyourself into Sweep And Prune, I believe this will be my goTo algorythm when thinking about this kind of problem, I believe one could optimize it to work with deletion and insertion.

Also the Idea of them being able to grab each other, could be interesting, at the beggining they would be able to simply have 4 hands one in each direction, they could have these active or inactive, BUT the calculation calculating the translation of a multinodenetwork with individual powers to it is not easy to compute.

Anyway in the end CTRNN is a theory, as I have found out, in the end it has to be broken down to D(discrete)TRNN, I am at the moment thinking about ways to actually make this system continuous, but there are people writing papers(not even 2 years old), that this is computable, so I believe I will not find possibilities to actually do this continuous.
http://web.mit.edu/~esontag/www/FTP_DIR/recur-survey-book.pdf
here is a link to a book that helped me alot.

For sh*ts and giggles I am working on a Quiditch(Harry Potter) simulator, where teams will be consistent of scryptbots, and I want to see how good these teams can get ;-). It is very promising. (2-Dimensional btw, but it seems to be scalable).