NeuralNet

table.duplicate = function(t)
	if(type(t) ~= "table") then return t end
	local t2 = {}
	for k,v in pairs(t) do
		if(type(v) == "table") then
			t2[k] = table.duplicate(v)
		else
			t2[k] = v
		end
	end
	return t2
end

-- Neural Networking API

local net = {}

function newInput()
	return {}
end
function newHidden(inputs, func)
	func = func or function(x) return 0 end
	local hidden = {}
	for i = 1, inputs, 1 do
		table.insert(hidden, func(i))
	end
	return hidden
end
function newOutput(inputs, func)
	func = func or function(x) return 0 end
	local output = {}
	for i = 1, inputs, 1 do
		table.insert(output, func(i))
	end
	return output
end


function newNet(...)
	local specs = {...}
	if(#specs < 2) then
		error("You need to specify at least how many inputs and outputs you want", 2)
	end
	for k,v in ipairs(specs) do
		if(type(v) ~= "number") then
			error("Argument #"..k..": Needs to be a number", 2)
		end
		if(v <= 0) then
			error("Argument #"..k..": You cannot have a zero-or-less node amount on a layer", 2)
		end
	end

	-- create net
	local net = {}

	-- adding input layer
	net[1] = {}
	for x = 1, specs[1], 1 do
		-- add individual input fields
		table.insert(net[1], newInput())
	end

	-- adding hidden layer
	for i = 2, #specs-1, 1 do
		-- create layer #net[i-1]
		net[i] = {}
		for x = 1, specs[i], 1 do
			--add individual neurons
			table.insert(net[i], newHidden(#net[i-1]))
		end
	end

	-- adding output layer
	net[#specs] = {}
	for x = 1, specs[#specs], 1 do
		--add individual neurons
		table.insert(net[#specs], newOutput(#net[#specs-1]))
	end
	return net
end
function runNet(net, inputs)
	if(type(net) ~= "table") then
		error("Arg #1: Net table expected, got "..type(net), 2)
	end
	if(type(inputs) ~= "table") then
		error("Arg #2: Input table expected, got "..type(inputs), 2)
	end
	if(#net[1] ~= #inputs) then
		error("The input sizes of net and inputs do not match", 2)
	end
	-- WORK check consistency

	-- Create instance to evaluate on without overriding weights
	local eval = table.duplicate(net)

	-- Insert inputs
	for i = 1, #inputs, 1 do
		eval[1][i] = inputs[i]
	end

	-- Evaluate hidden/output layers
	for x = 2, #eval, 1 do
		-- Insert evaluation for every neuron
		for i = 1, #eval[x], 1 do
			local nsum = 0
			for inp,weight in ipairs(eval[x][i]) do
				nsum = nsum + weight*eval[x-1][inp]
			end
			if(nsum >= 1) then -- threshold
				eval[x][i] = 1
			else
				eval[x][i] = 0
			end
		end
	end

	return eval[#eval]
end
function failrate(net, testset)
	--WORK
end
function backpropNet(net, testset, rounds)
	-- randomize testset order
	-- run complete testset and calculate the overall error
	-- adjust the weights
	-- check remaining time and loop
end