海纳百科

海纳百科

千行代码实现会玩《超级马里奥》的人工智能

吃瓜阿阳

友情提示点击顶部放大镜 可以使用站内搜索 记住我们的地址 www.hainabaike.com

诞生至今已经三十年的超级玛丽游戏,作为史上最畅销的电子游戏,游戏主人公马里奥的形象也为众人所熟知,尤其在七零后和八零后的童年有着不可磨灭的印记。然而全球著名的人工智能研究机构纷纷将目光重新放到这类经典的电视游戏中。

估值四亿美金的Deepmind

被Google以四亿美金收购的Deepmind团队演示其神奇的人工智能,同样一套系统可以在将近四十多款电视游戏中通过不断失措和自我学习,周而复始以致最终打出远远超出人类玩家的分数。

Deepmind神经网络
然而,能玩游戏的人工智能也不只Deepmind一家,今年年初,德国图宾根大学的认知模式小组在网上发布了一段视频,主角马里奥可以接受人类的语言教导,然后感知游戏周围环境并能自行做出正确游戏动作的决定,马里奥虽然能够理解英文语句和命令,并基于庞大的逻辑和语法树反馈回答人类究竟告诉过他什么,以及基于它所学知识做出正确的游戏动作。当与Deepmind相比,德国图宾根大学的人工智能仍需人类干预,自我试错和学习的能力弱了许多。

完全开源只有一千多行代码的AI

这不是关键,最为重要的是,后来有人开发出一套称之为MarI/O完全开源的AI,且代码居然只有一千多行!


MarI/O不像图宾根大学的程序,在它进入游戏甚至不知道游戏的终点是什么样的,相反,只是设定了几个简单的参数。这套AI有一个“Fitness”值,当马里奥向右移动时值增加,反之减少,因为它知道“Fitness”值是有好处的,一旦它发现向右移动数值能增加,这就会诱使它一直向右移动。

实际进化过程中,MarI/O并不会进行预测以改变其行动。通过进行不同的尝试,而不是做其“应该”做的事情,这样每次都会产生新的点子。当一个点子成功后,就会被记住,反之则被作废。就这样,超级马里奥在经历了34尝试后,完全通关了!当然,如果重新运行的话,这套AI机会肯定可以找到一条不同但不会更加成功的线路。

这种学习方式称之为神经网络进化拓扑结构(NeuroEvolution of Augmenting Topologies,简称NEAT),虽然这并不是一项新技术,但是在这里,作者却将其使用的非常高效。在一千多行Lua代码下,即实现了与估值四亿美金Deepmind类似的效果,不可不谓十分之神奇。当然,这仅仅是一个不错的演示,对于机器学习如果想要挑战一个更为强大的算法还有很长的路需要走。

对于想学习人工智能,将其算法或精神应用于现有的工作学习中,可以下载其源码本机模拟测试。

-- MarI/O by SethBling
-- Feel free to use this code, but please do not redistribute it.
-- Intended for use with the BizHawk emulator and Super Mario World or Super Mario Bros. ROM.

if gameinfo.getromname() == "Super Mario World (USA)" then
	Filename = "DP1.state"
	ButtonNames = {
		"A",
		"B",
		"X",
		"Y",
		"Up",
		"Down",
		"Left",
		"Right",
	}
elseif gameinfo.getromname() == "Super Mario Bros." then
	Filename = "SMB1-1.state"
	ButtonNames = {
		"A",
		"B",
		"Up",
		"Down",
		"Left",
		"Right",
	}
end

BoxRadius = 6
InputSize = (BoxRadius*2+1)*(BoxRadius*2+1)

Inputs = InputSize+1
Outputs = #ButtonNames

Population = 300
DeltaDisjoint = 2.0
DeltaWeights = 0.4
DeltaThreshold = 1.0

StaleSpecies = 15

MutateConnectionsChance = 0.25
PerturbChance = 0.90
CrossoverChance = 0.75
LinkMutationChance = 2.0
NodeMutationChance = 0.50
BiasMutationChance = 0.40
StepSize = 0.1
DisableMutationChance = 0.4
EnableMutationChance = 0.2

TimeoutConstant = 20

MaxNodes = 1000000

function getPositions()
	if gameinfo.getromname() == "Super Mario World (USA)" then
		marioX = memory.read_s16_le(0x94)
		marioY = memory.read_s16_le(0x96)

		local layer1x = memory.read_s16_le(0x1A);
		local layer1y = memory.read_s16_le(0x1C);

		screenX = marioX-layer1x
		screenY = marioY-layer1y
	elseif gameinfo.getromname() == "Super Mario Bros." then
		marioX = memory.readbyte(0x6D) * 0x100 + memory.readbyte(0x86)
		marioY = memory.readbyte(0x03B8)+16

		screenX = memory.readbyte(0x03AD)
		screenY = memory.readbyte(0x03B8)
	end
end

function getTile(dx, dy)
	if gameinfo.getromname() == "Super Mario World (USA)" then
		x = math.floor((marioX+dx+8)/16)
		y = math.floor((marioY+dy)/16)

		return memory.readbyte(0x1C800 + math.floor(x/0x10)*0x1B0 + y*0x10 + x%0x10)
	elseif gameinfo.getromname() == "Super Mario Bros." then
		local x = marioX + dx + 8
		local y = marioY + dy - 16
		local page = math.floor(x/256)%2

		local subx = math.floor((x%256)/16)
		local suby = math.floor((y - 32)/16)
		local addr = 0x500 + page*13*16+suby*16+subx

		if suby >= 13 or suby < 0 then
			return 0
		end

		if memory.readbyte(addr) ~= 0 then
			return 1
		else
			return 0
		end
	end
end

function getSprites()
	if gameinfo.getromname() == "Super Mario World (USA)" then
		local sprites = {}
		for slot=0,11 do
			local status = memory.readbyte(0x14C8+slot)
			if status ~= 0 then
				spritex = memory.readbyte(0xE4+slot) + memory.readbyte(0x14E0+slot)*256
				spritey = memory.readbyte(0xD8+slot) + memory.readbyte(0x14D4+slot)*256
				sprites[#sprites+1] = {["x"]=spritex, ["y"]=spritey}
			end
		end		

		return sprites
	elseif gameinfo.getromname() == "Super Mario Bros." then
		local sprites = {}
		for slot=0,4 do
			local enemy = memory.readbyte(0xF+slot)
			if enemy ~= 0 then
				local ex = memory.readbyte(0x6E + slot)*0x100 + memory.readbyte(0x87+slot)
				local ey = memory.readbyte(0xCF + slot)+24
				sprites[#sprites+1] = {["x"]=ex,["y"]=ey}
			end
		end

		return sprites
	end
end

function getExtendedSprites()
	if gameinfo.getromname() == "Super Mario World (USA)" then
		local extended = {}
		for slot=0,11 do
			local number = memory.readbyte(0x170B+slot)
			if number ~= 0 then
				spritex = memory.readbyte(0x171F+slot) + memory.readbyte(0x1733+slot)*256
				spritey = memory.readbyte(0x1715+slot) + memory.readbyte(0x1729+slot)*256
				extended[#extended+1] = {["x"]=spritex, ["y"]=spritey}
			end
		end		

		return extended
	elseif gameinfo.getromname() == "Super Mario Bros." then
		return {}
	end
end

function getInputs()
	getPositions()

	sprites = getSprites()
	extended = getExtendedSprites()

	local inputs = {}

	for dy=-BoxRadius*16,BoxRadius*16,16 do
		for dx=-BoxRadius*16,BoxRadius*16,16 do
			inputs[#inputs+1] = 0

			tile = getTile(dx, dy)
			if tile == 1 and marioY+dy < 0x1B0 then
				inputs[#inputs] = 1
			end

			for i = 1,#sprites do
				distx = math.abs(sprites[i]["x"] - (marioX+dx))
				disty = math.abs(sprites[i]["y"] - (marioY+dy))
				if distx <= 8 and disty <= 8 then
					inputs[#inputs] = -1
				end
			end

			for i = 1,#extended do
				distx = math.abs(extended[i]["x"] - (marioX+dx))
				disty = math.abs(extended[i]["y"] - (marioY+dy))
				if distx < 8 and disty < 8 then
					inputs[#inputs] = -1
				end
			end
		end
	end

	--mariovx = memory.read_s8(0x7B)
	--mariovy = memory.read_s8(0x7D)

	return inputs
end

function sigmoid(x)
	return 2/(1+math.exp(-4.9*x))-1
end

function newInnovation()
	pool.innovation = pool.innovation + 1
	return pool.innovation
end

function newPool()
	local pool = {}
	pool.species = {}
	pool.generation = 0
	pool.innovation = Outputs
	pool.currentSpecies = 1
	pool.currentGenome = 1
	pool.currentFrame = 0
	pool.maxFitness = 0

	return pool
end

function newSpecies()
	local species = {}
	species.topFitness = 0
	species.staleness = 0
	species.genomes = {}
	species.averageFitness = 0

	return species
end

function newGenome()
	local genome = {}
	genome.genes = {}
	genome.fitness = 0
	genome.adjustedFitness = 0
	genome.network = {}
	genome.maxneuron = 0
	genome.globalRank = 0
	genome.mutationRates = {}
	genome.mutationRates["connections"] = MutateConnectionsChance
	genome.mutationRates["link"] = LinkMutationChance
	genome.mutationRates["bias"] = BiasMutationChance
	genome.mutationRates["node"] = NodeMutationChance
	genome.mutationRates["enable"] = EnableMutationChance
	genome.mutationRates["disable"] = DisableMutationChance
	genome.mutationRates["step"] = StepSize

	return genome
end

function copyGenome(genome)
	local genome2 = newGenome()
	for g=1,#genome.genes do
		table.insert(genome2.genes, copyGene(genome.genes[g]))
	end
	genome2.maxneuron = genome.maxneuron
	genome2.mutationRates["connections"] = genome.mutationRates["connections"]
	genome2.mutationRates["link"] = genome.mutationRates["link"]
	genome2.mutationRates["bias"] = genome.mutationRates["bias"]
	genome2.mutationRates["node"] = genome.mutationRates["node"]
	genome2.mutationRates["enable"] = genome.mutationRates["enable"]
	genome2.mutationRates["disable"] = genome.mutationRates["disable"]

	return genome2
end

function basicGenome()
	local genome = newGenome()
	local innovation = 1

	genome.maxneuron = Inputs
	mutate(genome)

	return genome
end

function newGene()
	local gene = {}
	gene.into = 0
	gene.out = 0
	gene.weight = 0.0
	gene.enabled = true
	gene.innovation = 0

	return gene
end

function copyGene(gene)
	local gene2 = newGene()
	gene2.into = gene.into
	gene2.out = gene.out
	gene2.weight = gene.weight
	gene2.enabled = gene.enabled
	gene2.innovation = gene.innovation

	return gene2
end

function newNeuron()
	local neuron = {}
	neuron.incoming = {}
	neuron.value = 0.0

	return neuron
end

function generateNetwork(genome)
	local network = {}
	network.neurons = {}

	for i=1,Inputs do
		network.neurons[i] = newNeuron()
	end

	for o=1,Outputs do
		network.neurons[MaxNodes+o] = newNeuron()
	end

	table.sort(genome.genes, function (a,b)
		return (a.out < b.out)
	end)
	for i=1,#genome.genes do
		local gene = genome.genes[i]
		if gene.enabled then
			if network.neurons[gene.out] == nil then
				network.neurons[gene.out] = newNeuron()
			end
			local neuron = network.neurons[gene.out]
			table.insert(neuron.incoming, gene)
			if network.neurons[gene.into] == nil then
				network.neurons[gene.into] = newNeuron()
			end
		end
	end

	genome.network = network
end

function evaluateNetwork(network, inputs)
	table.insert(inputs, 1)
	if #inputs ~= Inputs then
		console.writeline("Incorrect number of neural network inputs.")
		return {}
	end

	for i=1,Inputs do
		network.neurons[i].value = inputs[i]
	end

	for _,neuron in pairs(network.neurons) do
		local sum = 0
		for j = 1,#neuron.incoming do
			local incoming = neuron.incoming[j]
			local other = network.neurons[incoming.into]
			sum = sum + incoming.weight * other.value
		end

		if #neuron.incoming > 0 then
			neuron.value = sigmoid(sum)
		end
	end

	local outputs = {}
	for o=1,Outputs do
		local button = "P1 " .. ButtonNames[o]
		if network.neurons[MaxNodes+o].value > 0 then
			outputs[button] = true
		else
			outputs[button] = false
		end
	end

	return outputs
end

function crossover(g1, g2)
	-- Make sure g1 is the higher fitness genome
	if g2.fitness > g1.fitness then
		tempg = g1
		g1 = g2
		g2 = tempg
	end

	local child = newGenome()

	local innovations2 = {}
	for i=1,#g2.genes do
		local gene = g2.genes[i]
		innovations2[gene.innovation] = gene
	end

	for i=1,#g1.genes do
		local gene1 = g1.genes[i]
		local gene2 = innovations2[gene1.innovation]
		if gene2 ~= nil and math.random(2) == 1 and gene2.enabled then
			table.insert(child.genes, copyGene(gene2))
		else
			table.insert(child.genes, copyGene(gene1))
		end
	end

	child.maxneuron = math.max(g1.maxneuron,g2.maxneuron)

	for mutation,rate in pairs(g1.mutationRates) do
		child.mutationRates[mutation] = rate
	end

	return child
end

function randomNeuron(genes, nonInput)
	local neurons = {}
	if not nonInput then
		for i=1,Inputs do
			neurons[i] = true
		end
	end
	for o=1,Outputs do
		neurons[MaxNodes+o] = true
	end
	for i=1,#genes do
		if (not nonInput) or genes[i].into > Inputs then
			neurons[genes[i].into] = true
		end
		if (not nonInput) or genes[i].out > Inputs then
			neurons[genes[i].out] = true
		end
	end

	local count = 0
	for _,_ in pairs(neurons) do
		count = count + 1
	end
	local n = math.random(1, count)

	for k,v in pairs(neurons) do
		n = n-1
		if n == 0 then
			return k
		end
	end

	return 0
end

function containsLink(genes, link)
	for i=1,#genes do
		local gene = genes[i]
		if gene.into == link.into and gene.out == link.out then
			return true
		end
	end
end

function pointMutate(genome)
	local step = genome.mutationRates["step"]

	for i=1,#genome.genes do
		local gene = genome.genes[i]
		if math.random() < PerturbChance then
			gene.weight = gene.weight + math.random() * step*2 - step
		else
			gene.weight = math.random()*4-2
		end
	end
end

function linkMutate(genome, forceBias)
	local neuron1 = randomNeuron(genome.genes, false)
	local neuron2 = randomNeuron(genome.genes, true)

	local newLink = newGene()
	if neuron1 <= Inputs and neuron2 <= Inputs then
		--Both input nodes
		return
	end
	if neuron2 <= Inputs then
		-- Swap output and input
		local temp = neuron1
		neuron1 = neuron2
		neuron2 = temp
	end

	newLink.into = neuron1
	newLink.out = neuron2
	if forceBias then
		newLink.into = Inputs
	end

	if containsLink(genome.genes, newLink) then
		return
	end
	newLink.innovation = newInnovation()
	newLink.weight = math.random()*4-2

	table.insert(genome.genes, newLink)
end

function nodeMutate(genome)
	if #genome.genes == 0 then
		return
	end

	genome.maxneuron = genome.maxneuron + 1

	local gene = genome.genes[math.random(1,#genome.genes)]
	if not gene.enabled then
		return
	end
	gene.enabled = false

	local gene1 = copyGene(gene)
	gene1.out = genome.maxneuron
	gene1.weight = 1.0
	gene1.innovation = newInnovation()
	gene1.enabled = true
	table.insert(genome.genes, gene1)

	local gene2 = copyGene(gene)
	gene2.into = genome.maxneuron
	gene2.innovation = newInnovation()
	gene2.enabled = true
	table.insert(genome.genes, gene2)
end

function enableDisableMutate(genome, enable)
	local candidates = {}
	for _,gene in pairs(genome.genes) do
		if gene.enabled == not enable then
			table.insert(candidates, gene)
		end
	end

	if #candidates == 0 then
		return
	end

	local gene = candidates[math.random(1,#candidates)]
	gene.enabled = not gene.enabled
end

function mutate(genome)
	for mutation,rate in pairs(genome.mutationRates) do
		if math.random(1,2) == 1 then
			genome.mutationRates[mutation] = 0.95*rate
		else
			genome.mutationRates[mutation] = 1.05263*rate
		end
	end

	if math.random() < genome.mutationRates["connections"] then
		pointMutate(genome)
	end

	local p = genome.mutationRates["link"]
	while p > 0 do
		if math.random() < p then
			linkMutate(genome, false)
		end
		p = p - 1
	end

	p = genome.mutationRates["bias"]
	while p > 0 do
		if math.random() < p then
			linkMutate(genome, true)
		end
		p = p - 1
	end

	p = genome.mutationRates["node"]
	while p > 0 do
		if math.random() < p then
			nodeMutate(genome)
		end
		p = p - 1
	end

	p = genome.mutationRates["enable"]
	while p > 0 do
		if math.random() < p then
			enableDisableMutate(genome, true)
		end
		p = p - 1
	end

	p = genome.mutationRates["disable"]
	while p > 0 do
		if math.random() < p then
			enableDisableMutate(genome, false)
		end
		p = p - 1
	end
end

function disjoint(genes1, genes2)
	local i1 = {}
	for i = 1,#genes1 do
		local gene = genes1[i]
		i1[gene.innovation] = true
	end

	local i2 = {}
	for i = 1,#genes2 do
		local gene = genes2[i]
		i2[gene.innovation] = true
	end

	local disjointGenes = 0
	for i = 1,#genes1 do
		local gene = genes1[i]
		if not i2[gene.innovation] then
			disjointGenes = disjointGenes+1
		end
	end

	for i = 1,#genes2 do
		local gene = genes2[i]
		if not i1[gene.innovation] then
			disjointGenes = disjointGenes+1
		end
	end

	local n = math.max(#genes1, #genes2)

	return disjointGenes / n
end

function weights(genes1, genes2)
	local i2 = {}
	for i = 1,#genes2 do
		local gene = genes2[i]
		i2[gene.innovation] = gene
	end

	local sum = 0
	local coincident = 0
	for i = 1,#genes1 do
		local gene = genes1[i]
		if i2[gene.innovation] ~= nil then
			local gene2 = i2[gene.innovation]
			sum = sum + math.abs(gene.weight - gene2.weight)
			coincident = coincident + 1
		end
	end

	return sum / coincident
end

function sameSpecies(genome1, genome2)
	local dd = DeltaDisjoint*disjoint(genome1.genes, genome2.genes)
	local dw = DeltaWeights*weights(genome1.genes, genome2.genes) 
	return dd + dw < DeltaThreshold
end

function rankGlobally()
	local global = {}
	for s = 1,#pool.species do
		local species = pool.species[s]
		for g = 1,#species.genomes do
			table.insert(global, species.genomes[g])
		end
	end
	table.sort(global, function (a,b)
		return (a.fitness < b.fitness)
	end)

	for g=1,#global do
		global[g].globalRank = g
	end
end

function calculateAverageFitness(species)
	local total = 0

	for g=1,#species.genomes do
		local genome = species.genomes[g]
		total = total + genome.globalRank
	end

	species.averageFitness = total / #species.genomes
end

function totalAverageFitness()
	local total = 0
	for s = 1,#pool.species do
		local species = pool.species[s]
		total = total + species.averageFitness
	end

	return total
end

function cullSpecies(cutToOne)
	for s = 1,#pool.species do
		local species = pool.species[s]

		table.sort(species.genomes, function (a,b)
			return (a.fitness > b.fitness)
		end)

		local remaining = math.ceil(#species.genomes/2)
		if cutToOne then
			remaining = 1
		end
		while #species.genomes > remaining do
			table.remove(species.genomes)
		end
	end
end

function breedChild(species)
	local child = {}
	if math.random() < CrossoverChance then
		g1 = species.genomes[math.random(1, #species.genomes)]
		g2 = species.genomes[math.random(1, #species.genomes)]
		child = crossover(g1, g2)
	else
		g = species.genomes[math.random(1, #species.genomes)]
		child = copyGenome(g)
	end

	mutate(child)

	return child
end

function removeStaleSpecies()
	local survived = {}

	for s = 1,#pool.species do
		local species = pool.species[s]

		table.sort(species.genomes, function (a,b)
			return (a.fitness > b.fitness)
		end)

		if species.genomes[1].fitness > species.topFitness then
			species.topFitness = species.genomes[1].fitness
			species.staleness = 0
		else
			species.staleness = species.staleness + 1
		end
		if species.staleness < StaleSpecies or species.topFitness >= pool.maxFitness then
			table.insert(survived, species)
		end
	end

	pool.species = survived
end

function removeWeakSpecies()
	local survived = {}

	local sum = totalAverageFitness()
	for s = 1,#pool.species do
		local species = pool.species[s]
		breed = math.floor(species.averageFitness / sum * Population)
		if breed >= 1 then
			table.insert(survived, species)
		end
	end

	pool.species = survived
end

function addToSpecies(child)
	local foundSpecies = false
	for s=1,#pool.species do
		local species = pool.species[s]
		if not foundSpecies and sameSpecies(child, species.genomes[1]) then
			table.insert(species.genomes, child)
			foundSpecies = true
		end
	end

	if not foundSpecies then
		local childSpecies = newSpecies()
		table.insert(childSpecies.genomes, child)
		table.insert(pool.species, childSpecies)
	end
end

function newGeneration()
	cullSpecies(false) -- Cull the bottom half of each species
	rankGlobally()
	removeStaleSpecies()
	rankGlobally()
	for s = 1,#pool.species do
		local species = pool.species[s]
		calculateAverageFitness(species)
	end
	removeWeakSpecies()
	local sum = totalAverageFitness()
	local children = {}
	for s = 1,#pool.species do
		local species = pool.species[s]
		breed = math.floor(species.averageFitness / sum * Population) - 1
		for i=1,breed do
			table.insert(children, breedChild(species))
		end
	end
	cullSpecies(true) -- Cull all but the top member of each species
	while #children + #pool.species < Population do
		local species = pool.species[math.random(1, #pool.species)]
		table.insert(children, breedChild(species))
	end
	for c=1,#children do
		local child = children
		addToSpecies(child)
	end

	pool.generation = pool.generation + 1

	writeFile("backup." .. pool.generation .. "." .. forms.gettext(saveLoadFile))
end

function initializePool()
	pool = newPool()

	for i=1,Population do
		basic = basicGenome()
		addToSpecies(basic)
	end

	initializeRun()
end

function clearJoypad()
	controller = {}
	for b = 1,#ButtonNames do
		controller["P1 " .. ButtonNames[b]] = false
	end
	joypad.set(controller)
end

function initializeRun()
	savestate.load(Filename);
	rightmost = 0
	pool.currentFrame = 0
	timeout = TimeoutConstant
	clearJoypad()

	local species = pool.species[pool.currentSpecies]
	local genome = species.genomes[pool.currentGenome]
	generateNetwork(genome)
	evaluateCurrent()
end

function evaluateCurrent()
	local species = pool.species[pool.currentSpecies]
	local genome = species.genomes[pool.currentGenome]

	inputs = getInputs()
	controller = evaluateNetwork(genome.network, inputs)

	if controller["P1 Left"] and controller["P1 Right"] then
		controller["P1 Left"] = false
		controller["P1 Right"] = false
	end
	if controller["P1 Up"] and controller["P1 Down"] then
		controller["P1 Up"] = false
		controller["P1 Down"] = false
	end

	joypad.set(controller)
end

if pool == nil then
	initializePool()
end

function nextGenome()
	pool.currentGenome = pool.currentGenome + 1
	if pool.currentGenome > #pool.species[pool.currentSpecies].genomes then
		pool.currentGenome = 1
		pool.currentSpecies = pool.currentSpecies+1
		if pool.currentSpecies > #pool.species then
			newGeneration()
			pool.currentSpecies = 1
		end
	end
end

function fitnessAlreadyMeasured()
	local species = pool.species[pool.currentSpecies]
	local genome = species.genomes[pool.currentGenome]

	return genome.fitness ~= 0
end

function displayGenome(genome)
	local network = genome.network
	local cells = {}
	local i = 1
	local cell = {}
	for dy=-BoxRadius,BoxRadius do
		for dx=-BoxRadius,BoxRadius do
			cell = {}
			cell.x = 50+5*dx
			cell.y = 70+5*dy
			cell.value = network.neurons[i].value
			cells[i] = cell
			i = i + 1
		end
	end
	local biasCell = {}
	biasCell.x = 80
	biasCell.y = 110
	biasCell.value = network.neurons[Inputs].value
	cells[Inputs] = biasCell

	for o = 1,Outputs do
		cell = {}
		cell.x = 220
		cell.y = 30 + 8 * o
		cell.value = network.neurons[MaxNodes + o].value
		cells[MaxNodes+o] = cell
		local color
		if cell.value > 0 then
			color = 0xFF0000FF
		else
			color = 0xFF000000
		end
		gui.drawText(223, 24+8*o, ButtonNames[o], color, 9)
	end

	for n,neuron in pairs(network.neurons) do
		cell = {}
		if n > Inputs and n <= MaxNodes then
			cell.x = 140
			cell.y = 40
			cell.value = neuron.value
			cells[n] = cell
		end
	end

	for n=1,4 do
		for _,gene in pairs(genome.genes) do
			if gene.enabled then
				local c1 = cells[gene.into]
				local c2 = cells[gene.out]
				if gene.into > Inputs and gene.into <= MaxNodes then
					c1.x = 0.75*c1.x + 0.25*c2.x
					if c1.x >= c2.x then
						c1.x = c1.x - 40
					end
					if c1.x < 90 then
						c1.x = 90
					end

					if c1.x > 220 then
						c1.x = 220
					end
					c1.y = 0.75*c1.y + 0.25*c2.y

				end
				if gene.out > Inputs and gene.out <= MaxNodes then
					c2.x = 0.25*c1.x + 0.75*c2.x
					if c1.x >= c2.x then
						c2.x = c2.x + 40
					end
					if c2.x < 90 then
						c2.x = 90
					end
					if c2.x > 220 then
						c2.x = 220
					end
					c2.y = 0.25*c1.y + 0.75*c2.y
				end
			end
		end
	end

	gui.drawBox(50-BoxRadius*5-3,70-BoxRadius*5-3,50+BoxRadius*5+2,70+BoxRadius*5+2,0xFF000000, 0x80808080)
	for n,cell in pairs(cells) do
		if n > Inputs or cell.value ~= 0 then
			local color = math.floor((cell.value+1)/2*256)
			if color > 255 then color = 255 end
			if color < 0 then color = 0 end
			local opacity = 0xFF000000
			if cell.value == 0 then
				opacity = 0x50000000
			end
			color = opacity + color*0x10000 + color*0x100 + color
			gui.drawBox(cell.x-2,cell.y-2,cell.x+2,cell.y+2,opacity,color)
		end
	end
	for _,gene in pairs(genome.genes) do
		if gene.enabled then
			local c1 = cells[gene.into]
			local c2 = cells[gene.out]
			local opacity = 0xA0000000
			if c1.value == 0 then
				opacity = 0x20000000
			end

			local color = 0x80-math.floor(math.abs(sigmoid(gene.weight))*0x80)
			if gene.weight > 0 then 
				color = opacity + 0x8000 + 0x10000*color
			else
				color = opacity + 0x800000 + 0x100*color
			end
			gui.drawLine(c1.x+1, c1.y, c2.x-3, c2.y, color)
		end
	end

	gui.drawBox(49,71,51,78,0x00000000,0x80FF0000)

	if forms.ischecked(showMutationRates) then
		local pos = 100
		for mutation,rate in pairs(genome.mutationRates) do
			gui.drawText(100, pos, mutation .. ": " .. rate, 0xFF000000, 10)
			pos = pos + 8
		end
	end
end

function writeFile(filename)
        local file = io.open(filename, "w")
	file:write(pool.generation .. "
")
	file:write(pool.maxFitness .. "
")
	file:write(#pool.species .. "
")
        for n,species in pairs(pool.species) do
		file:write(species.topFitness .. "
")
		file:write(species.staleness .. "
")
		file:write(#species.genomes .. "
")
		for m,genome in pairs(species.genomes) do
			file:write(genome.fitness .. "
")
			file:write(genome.maxneuron .. "
")
			for mutation,rate in pairs(genome.mutationRates) do
				file:write(mutation .. "
")
				file:write(rate .. "
")
			end
			file:write("done
")

			file:write(#genome.genes .. "
")
			for l,gene in pairs(genome.genes) do
				file:write(gene.into .. " ")
				file:write(gene.out .. " ")
				file:write(gene.weight .. " ")
				file:write(gene.innovation .. " ")
				if(gene.enabled) then
					file:write("1
")
				else
					file:write("0
")
				end
			end
		end
        end
        file:close()
end

function savePool()
	local filename = forms.gettext(saveLoadFile)
	writeFile(filename)
end

function loadFile(filename)
        local file = io.open(filename, "r")
	pool = newPool()
	pool.generation = file:read("*number")
	pool.maxFitness = file:read("*number")
	forms.settext(maxFitnessLabel, "Max Fitness: " .. math.floor(pool.maxFitness))
        local numSpecies = file:read("*number")
        for s=1,numSpecies do
		local species = newSpecies()
		table.insert(pool.species, species)
		species.topFitness = file:read("*number")
		species.staleness = file:read("*number")
		local numGenomes = file:read("*number")
		for g=1,numGenomes do
			local genome = newGenome()
			table.insert(species.genomes, genome)
			genome.fitness = file:read("*number")
			genome.maxneuron = file:read("*number")
			local line = file:read("*line")
			while line ~= "done" do
				genome.mutationRates[line] = file:read("*number")
				line = file:read("*line")
			end
			local numGenes = file:read("*number")
			for n=1,numGenes do
				local gene = newGene()
				table.insert(genome.genes, gene)
				local enabled
				gene.into, gene.out, gene.weight, gene.innovation, enabled = file:read("*number", "*number", "*number", "*number", "*number")
				if enabled == 0 then
					gene.enabled = false
				else
					gene.enabled = true
				end

			end
		end
	end
        file:close()

	while fitnessAlreadyMeasured() do
		nextGenome()
	end
	initializeRun()
	pool.currentFrame = pool.currentFrame + 1
end

function loadPool()
	local filename = forms.gettext(saveLoadFile)
	loadFile(filename)
end

function playTop()
	local maxfitness = 0
	local maxs, maxg
	for s,species in pairs(pool.species) do
		for g,genome in pairs(species.genomes) do
			if genome.fitness > maxfitness then
				maxfitness = genome.fitness
				maxs = s
				maxg = g
			end
		end
	end

	pool.currentSpecies = maxs
	pool.currentGenome = maxg
	pool.maxFitness = maxfitness
	forms.settext(maxFitnessLabel, "Max Fitness: " .. math.floor(pool.maxFitness))
	initializeRun()
	pool.currentFrame = pool.currentFrame + 1
	return
end

function onExit()
	forms.destroy(form)
end

writeFile("temp.pool")

event.onexit(onExit)

form = forms.newform(200, 260, "Fitness")
maxFitnessLabel = forms.label(form, "Max Fitness: " .. math.floor(pool.maxFitness), 5, 8)
showNetwork = forms.checkbox(form, "Show Map", 5, 30)
showMutationRates = forms.checkbox(form, "Show M-Rates", 5, 52)
restartButton = forms.button(form, "Restart", initializePool, 5, 77)
saveButton = forms.button(form, "Save", savePool, 5, 102)
loadButton = forms.button(form, "Load", loadPool, 80, 102)
saveLoadFile = forms.textbox(form, Filename .. ".pool", 170, 25, nil, 5, 148)
saveLoadLabel = forms.label(form, "Save/Load:", 5, 129)
playTopButton = forms.button(form, "Play Top", playTop, 5, 170)
hideBanner = forms.checkbox(form, "Hide Banner", 5, 190)

while true do
	local backgroundColor = 0xD0FFFFFF
	if not forms.ischecked(hideBanner) then
		gui.drawBox(0, 0, 300, 26, backgroundColor, backgroundColor)
	end

	local species = pool.species[pool.currentSpecies]
	local genome = species.genomes[pool.currentGenome]

	if forms.ischecked(showNetwork) then
		displayGenome(genome)
	end

	if pool.currentFrame%5 == 0 then
		evaluateCurrent()
	end

	joypad.set(controller)

	getPositions()
	if marioX > rightmost then
		rightmost = marioX
		timeout = TimeoutConstant
	end

	timeout = timeout - 1

	local timeoutBonus = pool.currentFrame / 4
	if timeout + timeoutBonus <= 0 then
		local fitness = rightmost - pool.currentFrame / 2
		if gameinfo.getromname() == "Super Mario World (USA)" and rightmost > 4816 then
			fitness = fitness + 1000
		end
		if gameinfo.getromname() == "Super Mario Bros." and rightmost > 3186 then
			fitness = fitness + 1000
		end
		if fitness == 0 then
			fitness = -1
		end
		genome.fitness = fitness

		if fitness > pool.maxFitness then
			pool.maxFitness = fitness
			forms.settext(maxFitnessLabel, "Max Fitness: " .. math.floor(pool.maxFitness))
			writeFile("backup." .. pool.generation .. "." .. forms.gettext(saveLoadFile))
		end

		console.writeline("Gen " .. pool.generation .. " species " .. pool.currentSpecies .. " genome " .. pool.currentGenome .. " fitness: " .. fitness)
		pool.currentSpecies = 1
		pool.currentGenome = 1
		while fitnessAlreadyMeasured() do
			nextGenome()
		end
		initializeRun()
	end

	local measured = 0
	local total = 0
	for _,species in pairs(pool.species) do
		for _,genome in pairs(species.genomes) do
			total = total + 1
			if genome.fitness ~= 0 then
				measured = measured + 1
			end
		end
	end
	if not forms.ischecked(hideBanner) then
		gui.drawText(0, 0, "Gen " .. pool.generation .. " species " .. pool.currentSpecies .. " genome " .. pool.currentGenome .. " (" .. math.floor(measured/total*100) .. "%)", 0xFF000000, 11)
		gui.drawText(0, 12, "Fitness: " .. math.floor(rightmost - (pool.currentFrame) / 2 - (timeout + timeoutBonus)*2/3), 0xFF000000, 11)
		gui.drawText(100, 12, "Max Fitness: " .. math.floor(pool.maxFitness), 0xFF000000, 11)
	end

	pool.currentFrame = pool.currentFrame + 1

	emu.frameadvance();
end

    标签:

    发布评论 条评论)

    评论列表