The Powder Toy

peeps

Hello guys

I am currently developing an experimental neural network based trainable element system.

What i want to do:

an API for dynamic lua element generation and training:

you create a special particle, or specify the particular data, then you specify the state of simulation around that particle, and what the particle itself and the stuff arount that particle would be look like after a tick.

you create as many examples of the desired interactions that you want to create for this element, using a "before" and an "after" format.

Then the neural network should be able to "learn" what it should do next, and after a few hundred training samples, should behave like what you want it to be. the particle basically learn it's proper update function through machine learning.

these are the samples of my code that i am currently working on.

Will this work? do you have any suggestions? for example, to improve this code or to submit some functions, so i can bring this into existence? thanks.

function LinRec_AC(level,offset,sensitivity)
x=(level-offset)*sensitivity
if x>0 then
return x
else
return x*0.1
end
end

function LinRec_BP(output,offset,sensitivity)
if output>0 then
return (output/sensitivity)+offset
else
return (output*10/sensitivity)+offset
end
end

function FullyConnected.initialize_neurons(number,inter_mediate_depth,i_o_range)
mapping={}
Cv=1
mapping[Cv]=FullyConnected.initialize_input(i_o_range)
Cv=Cv+1
for i=1,inter_mediate_depth,1 do
  mapping[Cv]=FullyConnected.initialize_layer(number)
  Cv=Cv+1
end
mapping[Cv]=FullyConnected.initialize_output(i_o_range)
for j=1,Cv-1,1 do
  FullyConnected.initialize_connection(mapping[j],mapping[j+1])
end
return mapping
end

function FullyConnected.LoadFromElement(mapping,x,y,Mem)
bits=bit.tobit(tpt.get_property("tmp",x,y))
for index,Neui in pairs(mapping[1]) do
  if Neui.typ==input then
   part_type=tpt.get_property("type",x+Neui.x,y+Neui.y)
   if Neui.typ=="typeU2b" then
    Neui.level=math.floor(part_type/64)*Neui.multiplier
   elseif Neui.typ=="typeM3b" then
    Neui.level=math.floor(part_type%64/8)*Neui.multiplier
   elseif Neui.typ=="typeL3b" then
    Neui.level=part_type%8*Neui.multiplier
   else
    Neui.level=tpt.get_property(Neui.property,x+Neui.x,y+Neui.y)*Neui.multiplier
   end
  elseif Neui.typ=mem_input then
   bmask=bit.lshift(bit.tobit(1),Neui.place-1)
   bres=bit.rshift(bit.band(bits,bmask),Neui.place-1)
   Neui.level=bres*Neui.multiplier
  end
end
end

also, is this neuron design adequate?

default_Neuron={typ=default,level=0,offset=0.5,sensitivity=1,Weights_from={amount=0},Weights_to={amount=0}}
input_neuron={typ=input,multiplier=0.1,offset=0.5,sensitivity=1,Weights_to={amount=0},level=0,x=0,y=0,property="life"}
input_memory_neuron={typ=mem_input,multiplier=1,offset=0.5,sensitivity=1,Weights_to={amount=0},level=0,place=1}
output_neuron={typ=output,multiplier=10,offset=0.5,sensitivity=1,Weights_from={amount=0},level=0,x=0,y=0,property="life"}
output_memory_neuron={typ=mem_output,threshold=0.5,offset=0.5,sensitivity=1,Weights_from={amount=0},level=0,place=1}

are there any improvement for this?

This is just pretty much a WIP. I am welcome for any one that could help.

Update: the full WIP code is behind this link:

https://pastebin.ubuntu.com/p/hCgRsyk5tR/

NF

Looks nice, I wish I could code but unfortunately lack the patience. Wish you luck.

alice_loona_OT12

This is nice code, but you forgot to use == instead of = on lines 346 and 381.

Edit: Sorry for necroposting, but I had to comment to alert OP of the mistake.

regk

can you make a regular pastebin link and not a ubuntu pastebin link please?

Trainable elements?