Recipe: neuroNet104findAminosOnly3PUBG2.Lua

created by zo3xiaJonWeinberg

Profile

• Name: neuroNet104findAminosOnly3PUBG2.Lua
• ID: 105130
• Shared with: Public
• Parent: None
• Children:
  • neuroNet105findAminosHotFireBorder.Lua
• Created on: August 22, 2021 at 19:33 PM UTC
• Updated on: August 22, 2021 at 19:33 PM UTC
• Description:

  Login to Fold.it to download. ; Weile fang zai ni de mingdan, ; xian denglu zai fold.it .

  If you see Error: Can't find specified ; recipe, then it is currently unshared. ; But you can ask for it.

  click Show Output terminal ; button. ; open Behavior tab see fitness ; rising and ; Check Recipe Score Modding. ; keep info width within 1 line ; for hover info. ; This neural network A.i. is a work ; in progress. ; Welcome to rate or expand the ; hidden layers or add inputs, ; outputs like parameters ; or teach it to count. ; An earlier ReLU version already ; did somewhat OK ; on the XOR problem, though it just ; converged its output from 80 to 0 ; for both true 1 and false 0.. ; Maybe try ReLU for hiddens only. ; Best for new puzzles like Linker ; people are unfamiliar with, ; or simple pred. puzzles. ; Works for multistart arabidopsis.

  mar 12 public neuroNet41- ; differentiateNegatives ; I ought to write these ; updates in chronological ; order (instead of ; the top being ; the newest).

  mar6 accidentally modified old version, had to merge edits ; feb.26 alphago likes the "kicks". ; count the double kicks.

  feb 23. this NN finally got a higher ; score than AFK3 on beg. cov. ; but it had 23 cuts. ; shifu: ; layer , neuron: ; dendrites: -0.1, -0.9, 0.4, 0.1, , ; bias: 0.6, ; neuron: ; dendrites: 0.2, -1, -0.4, 0.3, , ; bias: 1 ; layer , neuron: ; dendrites: 0.2, 0.9, , bias: -0.9

  So now that the NN's are proven ; ready, I'm making it harder ; by making cuts or filters disabled ; only give half or 0 points.

  Feb. 9, more comments, ; and save shifu NN as note ; which can be copied by user. ; example: ; shifu: ; layer , neuron: ; dendrites: -0.1, -0.9, 0.4, 0.1, , ; bias: 0.6, ; neuron: ; dendrites: 0.2, -1, -0.4, 0.3, , ; bias: 1 ; layer , neuron: ; dendrites: 0.2, 0.9, , bias: -0.9

  publishing. ; quick fix of bug where pokemon 1 ; could not be recorded as strongest ; or 2nd strongest. ; Feb. 7. noticed a.i. using time ; consuming action MSG All ; perhaps to bide time ; because I did deduct negative ; score somewhat (alpha=0.1) ; but divided it by time. ; Update feb. 2021, fixed index out ; of bounds for insert.

  Jan. 2021 ; Fixed Design bug del. res ; out of bounds (del. 2x). ; Still has out of bounds insert bug. ; Discarding update ideal14's ; advanced evolution in favor of ; "disciple exceeds master". ; Also trying out muta worst seg.

  2020 ; It likes Rebuild worst for prediction. ; Fixed choosing best scorer ; subtle bug. ; Had been assigning non-weakest ; as strongest, so last critter tended ; to be "strongest". ; Also added runner-up. ; now without creditbest. ; and, or? ; greyscale and teacher ; w/o random, outputs converge to 1 ; after waking from a 9 hour sleep ; Now with random inputs. ; W/ random inputs, ; convergence after 1 hour (hora) ; realized the xor function was bugged ; fixed xor, changed stuff ; converges to 98.5% fitness ; after 10 min. ; "Code 871". ; now IRL problem, clash. ; middle ground 0.9 after 5 mins

  2021-04-02 ; "I don't always mutate the dendrite, ; but when I do, it's by up to 0.5"

  4-24 ; 3b->4

  2021-4-24 9AM Data balance ; because stochastic ; gradient descent ; is more complicated

  '21-5-21 ; Bifurcated axon for more ; realistic simulation of neurons. ; And re-added a hidden layer. ; But these make solving XOR slower. ; Check previous versions for ; a good XOR demo.

  road less traveled test for forked ; axons, which depend on ; neighboring electrical activity. ; --_- ; __-_

  Jun 11. ; Could I draw a NN by hand that ; tries every number within a range ; an equal number of times? ; Could it count? If the input was not ; run through a function, maybe ; I could. If it was squeezed with ; tanh, I doubt it. ; So it seems only fair to give the ; unsqueezed input to the A.i.

  V.87: Regret retractive braces

  June 2X: Summer, don't wanna ; run it myself. Too dry.

  Jul. 16: ; Peanut and [Western only] ; eggplant are shaped like oxalate, ; and olivetol is shaped ; like the "green stuff" ; (and some lichen leaves too, ; I guess), ; and this neural network is ; shaped more like gnocchi now, ; because I dreamt of gnocchi ; and it's the NN beginner shape.

  Aug. 22 ; I was having trouble restricting ; the NN outputs to the range ; of acceptable "moves", ; so I made a penalty proportionate ; to the distance out of bounds, ; hence the PUBG ring of ; fire mechanic. ; Like, you can't set an amino acid to ; be DNA or vinegar, so the NN ; needs to be punished. ; Hotfix to stop NN from being ; recurrent, ; as I'm probably doing it wrong. ; Noon hotfix 2 to further avoid ; recurrence by ignoring last ; action paramater.

Code

--keep blank line at top 4 debug
help()
--Fn Down or Up to jump, but creates "," at top

humble=true --humble=no clover axon
testXor=true
hiddenLayers=3
yellowWood=1 --math.random(
chooseAminos=true--false--
debugActions=false

askXor=dialog.CreateDialog("Test XOR function?")
askXor.l1=dialog.AddLabel("Exclusive Or function")
askXor.l2=dialog.AddLabel("Is Bella in an exclusive ")
askXor.l3=dialog.AddLabel("relationship with Edward or Jacob ")
askXor.l4=dialog.AddLabel("but not both?")
askXor.play=dialog.AddButton("PlayFoldit",0)

askXor.test=dialog.AddButton("TestXOR",1)
resAskXor=dialog.Show(askXor)
if resAskXor==0 then
testXor=false


--Josh fuse snippet
behavior.SetHidingImportance(0.5)
five=behavior.GetHidingImportance()
behavior.SetHidingImportance(0.6)
six=behavior.GetHidingImportance()
if five==six then 
local ask = dialog.CreateDialog("Recipe modding is disabled")
ask.l1=dialog.AddLabel("You may want to")-- cancel,")
ask.l2=dialog.AddLabel(" open(dakai) the Behavior/Comporte tab,")
ask.l3=dialog.AddLabel(" and click Recipe Score Modding.")--, then re-run")
ask.l4=dialog.AddLabel("You don't need to cancel")
ask.l5=dialog.AddLabel("buyong quxiao")

ask.OK = dialog.AddButton("OK/Haode",1) --Wo bu guan", 1)

ask.Cancel = dialog.AddButton("Cancel/quxiao", 0)

res=dialog.Show(ask)
if res==0 then exitexitexit() end
end
end
function resetBehavior()

    behavior.SetClashImportance ( 1.0 )

    behavior.SetSidechainHBondImportance ( 1.0 )

    behavior.SetBackboneHBondImportance ( 1.0 )

    behavior.SetPackingImportance ( 1.0 )

behavior.SetHidingImportance ( 1.0 )
    behavior.SetPairwiseImportance ( 1.0 ) 
    behavior.SetDensityImportance ( 1.0 )--jon 
end

--end Josh's Fuse snippet

worstPossible=-999999.99999
multiStart=false
HASMUTABLE=false
descrTxt=puzzle.GetDescription()
if #descrTxt>0 and (descrTxt:find("design") or descrTxt:find("designs")) then        
HASMUTABLE=true        IDEALCHECK=true        HANDFOLD=true    end
--copied from AFK3.1.1 using new foldit editor

if #descrTxt>0 and (descrTxt:find("multi-start") or descrTxt:find("start positions")) then 
    multiStart=true
end

--if HASMUTABLE then environInput1=1 else environInput1=0 end
print("muta ")
print(HASMUTABLE) --can't cat 

-----
--So I cut and pasted these functions from BWP here for AFK.
NonLockedSegList={}
function segCnt() return structure.GetCount() end --jon
function IsUnlockedSeg( seg1 )
   return not structure.IsLocked( seg1 )
end
function FindAllMovableSegs( )
    --segCnt=structure.GetCount() --jon lazy
    NonLockedSegList={} --jon reinit in case deleting res.
    for i=1, structure.GetCount() do --jon
        if structure.IsLocked( i ) then 
            PuzzleHasLockedSegs = true
        else
            NonLockedSegList[ #NonLockedSegList + 1] = i
        end
    end
    if PuzzleHasLockedSegs then
        for i=1, structure.GetCount() do --jon
            if IsUnlockedSeg( i ) then 
                MinUnlockedSeg = i
                break
            end
        end
        for i=segCnt(), 1, -1 do
            if IsUnlockedSeg( i ) then 
                MaxUnlockedSeg = i
                break
            end
        end
    end
    return false
end
FindAllMovableSegs()
------

param=099
squeezeToInt=function(max)
    --param anything like -99.8
        -- tanh -> -1 to 1
        -- 0 to 2
        -- 0 to 1
        -- 0 to max-1
        -- 1 to max+.99 for rounding fairness
        --choice=math.floor((math.tanh(param)+1)/2*(max-1)+1.99) 
        min=1
        choice=math.floor(param%(max-min)+min)
        --Lua makes squeezing into range lil. complicated because lists start at 1
    return choice
end
squeezeFloat=function(min,max) --bug where output constant value at max
    --param anything like -99.8
        -- tanh -> -1 to 1
        -- 0 to 2
        -- 0 to 1
        -- 0 to max-1
        -- 1 to max+.99 for rounding fairness
        --choice=((math.tanh(param)+1)/2*((max-min)-1)+1)+min 
        --Lua makes squeezing into range lil. complicated because lists start at 1
    --choice=eulerAct(param)*(max-min)+min 
    choice=param%(max-min)+min
    print("choice is "..choice)
    return choice
end
eulerAct=function(a)
    --from xviniette's AsteroidsLearning Neuroevolution.js on Github
    --activation:function(a){
    ap = (-a)/1
    return (1/(1 + math.exp(ap))) --like tanh but from 0 to 1 instead of -1 to 1
    --exp(x) is Euler's constant e^x
end
print("euler "..eulerAct(-4.5))


------------A.i. learns v7 filters
performAction={}
actionDescription={}
function printTimeAndScore(startTime, scoreGain, currentStateString)
    local ss = (os.time() - startTime) % 60
    local mm = (((os.time() - startTime - ss) % 3600) / 60)
    local hh = (os.time() - startTime - mm*60 - ss) / 3600
    print("Time: " .. string.format("%02d", hh) .. ":" .. string.format("%02d", mm) .. ":" .. 
string.format("%02d", ss) .. ". " .. currentStateString)
    print(" Score: " .. current.GetEnergyScore() .. 
", Total: +" .. scoreGain)
end
bookmark=0
function initializeActions()
    performAction = {}
    actionDescription = {}

    --
    performAction[0] = function()
        local ss=math.random(1,structure.GetCount()-3)
        selection.SelectRange(ss,ss+3)
        structure.MutateSidechainsSelected(1) --All(1)
    end
    actionDescription[0] = "Mutate(1)"
    --

    table.insert(performAction,function() lastOutput=0 --brainOutput=1 
        param=0
end)
    actionDescription[1] = "Reset prev. brainOutput to not spin out of control"
    --keep this at first action, because Lua makes 3.56e+89 % 5 or % 6 to 1 I mean 0

    performAction[2] = function()
        local startSegmentNumber = math.random(1, structure.GetCount() - 9)
        local endSegmentNumber = math.random(startSegmentNumber + 8, structure.GetCount())
        local b=band.AddBetweenSegments(startSegmentNumber, endSegmentNumber)
        band.SetStrength(b, 10) --jon b
        band.SetGoalLength(b, structure.GetDistance(startSegmentNumber, endSegmentNumber) * 0.9)
    end
    actionDescription[2] = "Add random band"
    performAction[3] = function()
        band.DeleteAll()
    end
    actionDescription[3] = "Delete all bands"

    recentbest.Save() --overwrite it first
    performAction[4] = function()
        recentbest.Restore() --still somehow accessing
        FindAllMovableSegs()
        --print("banned")
    end
    actionDescription[4] = "Restore recent best"

    performAction[5] = function()
        local startSegmentNumber = math.random(1,structure.GetCount() - 3)
        selection.DeselectAll()
        selection.SelectRange(startSegmentNumber, startSegmentNumber + 3)
        structure.RebuildSelected(2)
        selection.DeselectAll()
    end
    actionDescription[5] = "Rebuild random segment"
    performAction[6] = function()
        behavior.SetClashImportance(1)
    end
    actionDescription[6] = "Set clash importance to 1"
    performAction[7] = function()
        behavior.SetClashImportance(0.5)
    end
    actionDescription[7] = "Set clash importance to 0.5"
    performAction[8] = function()
        behavior.SetClashImportance(0.02)
    end
    actionDescription[8] = "Set clash importance to 0.02"
    performAction[9] = function()
        local startSegmentNumber = math.random(1, structure.GetCount() - 3)
        selection.DeselectAll()
        selection.SelectRange(startSegmentNumber, startSegmentNumber + 3)
        structure.IdealizeSelected()
        selection.DeselectAll()
    end
    table.insert(actionDescription,"Idealize random segment")

    
    performAction[10] = function()
        FindAllMovableSegs() 
        cut=(NonLockedSegList[math.random(#NonLockedSegList) ])
        structure.InsertCut(cut)
        structure.WiggleAll(1)
        structure.DeleteCut(cut)
    end
    actionDescription[10] = "Wiggle(1 band safe 2 sided interf." --interface --no red lines

    --table.insert(performAction,function() structure.MutateSidechainsAll(1) end)
    --table.insert(actionDescription,"mutate1 moved from action 0")


    function ParseError ( errmsg )
        local reason
        local start, stop, line, msg
        start, stop, line, msg = errmsg:find ( ":(%d+):%s()" )
        if msg ~= nil then
            errmsg = errmsg:sub ( msg, #errmsg )
        end
        return errmsg    
    end

--
--  SafeSnap uses the Lua protected call function, pcall,
--  to call rotamer.SetRotamer
--
--  SafeSnap returns a numeric return code and an error message.
--
--  The return codes are:
--
--   0 - successful, error message is nil
--  -1 - bad rotamer index
--
--  Errors other than a bad rotamer index are rethrown using 
--  the Lua function error().
--  
    function SafeSnap ( segIdx, rotIdx )
    --
    --  error message when attempting to snap to an invalid rotamer
    --
        local BADSNAP = "(snap index out of bounds)" 
    --
    --  don't allow a bad rotamer to stop us
    --
        local good, errmsg = pcall ( rotamer.SetRotamer, segIdx, rotIdx )
        if good then
            return 0, nil
        else
            local err2 = ParseError ( errmsg )
            local errp = err2:find ( BADSNAP )
            if errp ~= nil then
                return -1, err2
            end
            error ( errmsg, 0 )
        end
    end


    table.insert(performAction, function() SafeSnap(math.random(structure.GetCount()), param) end)
    table.insert(actionDescription,"pcall rotamer for ligand or res")
    
    table.insert(performAction,function()
        local segmentIndex1=math.random(structure.GetCount())
        local segmentIndex2=math.random(structure.GetCount())
        local atomIndex1=math.random(structure.GetAtomCount(segmentIndex1))
        
        local atomIndex2=math.random(structure.GetAtomCount(segmentIndex2))
        
        local symnr2=math.random(0,structure.GetSymCount())
        return band.AddBetweenSegments(segmentIndex1,segmentIndex2,atomIndex1,atomIndex2,symnr2)
    end)
    table.insert(actionDescription,
    "Add a band between the two segments. By default bands to the backbone. Specify an atom number from 0 to structure.GetAtomCount() to band to a different atom. (0 for the default backbone atom) Returns band number. Specify an symnr from 0 t")

    table.insert(performAction,function()
        local lengthOfSegment = math.random(1, 3)
        local segmentInformation = {}
        local workSegmentNumber = math.random(1, 3)
        local function sortSegments(segmentInformation)
            for i = 1, #segmentInformation - 1 do
                for j = i + 1, #segmentInformation do
                    if segmentInformation[i][3] > segmentInformation[j][3] then
                        segmentInformation[i], segmentInformation[j] = segmentInformation[j], 
segmentInformation[i]
                    end
                end
            end
            return segmentInformation
        end
        for i = 1, structure.GetCount() - lengthOfSegment + 1 do
            segmentInformation[i] = {i, i + lengthOfSegment - 1, current.GetSegmentEnergyScore(i) }
            for j = 2, lengthOfSegment do
                segmentInformation[i][3] = segmentInformation[i][3] + current.GetSegmentEnergyScore(i 
+ j - 1)
            end
        end
        segmentInformation = sortSegments(segmentInformation)
        selection.DeselectAll()
        selection.SelectRange(math.max(segmentInformation[workSegmentNumber][1] - 1, 1), 
math.min(segmentInformation[workSegmentNumber][2] + 1, structure.GetCount()))
        structure.RebuildSelected(2)
        selection.DeselectAll()
    end)
    table.insert(actionDescription,"Rebuild worst segment")

    table.insert(performAction,function()
        local startSegmentNumber = math.random(1, structure.GetCount() - 9)
        local endSegmentNumber = math.random(startSegmentNumber + 8, structure.GetCount())
        local b=band.AddBetweenSegments(startSegmentNumber, endSegmentNumber)
        band.SetStrength(b, 10)
        band.SetGoalLength(b, structure.GetDistance(startSegmentNumber, endSegmentNumber) * 1.1)
    end)
    table.insert(actionDescription,"expand band")

    table.insert(performAction,function()
        copier=math.random(1,structure.GetCount())
        structure.SetSecondaryStructure(copier,structure.GetSecondaryStructure(
                math.max(math.min(copier+math.random(0,1)*2-1,1),structure.GetCount())
            )
        )
    end)
    table.insert(actionDescription,"extend ss")

    table.insert(performAction,function()       
        freeze.Freeze(math.random(1,structure.GetCount()),false,true) --bb --sc
    end)
    table.insert(actionDescription,"freeze sc")

    table.insert(performAction,function()       
        freeze.Freeze(math.random(1,structure.GetCount()),true,false) --bb --sc
    end)
    table.insert(actionDescription,"freeze bb")

    table.insert(performAction,function()       
        freeze.Freeze(math.random(1,structure.GetCount()),true,true) --bb --sc
    end)
    table.insert(actionDescription,"freeze bb sc") --threoretically could freeze neither for testing

    table.insert(performAction,function()       
        freeze.Unfreeze(math.random(1,structure.GetCount()),false,true) --bb --sc
    end)
    table.insert(actionDescription,"unfreeze sc")

    table.insert(performAction,function()       
        freeze.Unfreeze(math.random(1,structure.GetCount()),true,false) --bb --sc
    end)
    table.insert(actionDescription,"unfreeze bb")


    act="freeze.Unfreeze(math.random(1,structure.GetCount()),true,true) --bb --sc"
    table.insert(performAction,function()       
        pcall(act)
        print(act)
    end)
    table.insert(actionDescription,act)--"unfreeze bb")
    act="kokokokok" --still works, at least in Lua
    
    table.insert(performAction,function()       
        freeze.UnfreezeAll() --bb --sc
    end)
    table.insert(actionDescription,"unfreeze all")

    
    table.insert(performAction,function()       
        freeze.FreezeAll() --bb --sc
    end)
    table.insert(actionDescription,"freeze all")

    table.insert(performAction,function() behavior.SetBackboneHBondImportance(math.random(0,300)/100)
    end)
    table.insert(actionDescription,"behavior.SetBackboneHBondImportance (rand)")

    table.insert(performAction,function() behavior.SetBackboneHBondImportance(1)
    end)
    table.insert(actionDescription,"behavior.SetBackboneHBondImportance(1)")

    table.insert(performAction,function() resetBehavior() end)
    table.insert(actionDescription,"resetBehavior")

    table.insert(performAction,function()
        behavior.SetFiltersDisabled(true)
    end)
    table.insert(actionDescription,"disable filters")

    table.insert(performAction,function()
        behavior.SetFiltersDisabled(false)
    end)
    table.insert(actionDescription,"enable filters")
    
    table.insert(performAction,function()
        behavior.SetDensityImportance(squeezeFloat(0,3))
    end)
    table.insert(actionDescription,"density, enable recipe score modding first")

    table.insert(performAction,function()
        behavior.SetDensityImportance(1)
    end)
    table.insert(actionDescription,"density1, enable recipe score modding first")

    table.insert(performAction,function()
        options={"l","m","a"}
        if behavior.HighPowerAllowed() then table.insert(options,"h") end
        behavior.SetWigglePower(options[squeezeToInt(#options)])
        --old buggy algo: math.floor(#options/2*(math.tanh(param)+1))
    end)
    table.insert(actionDescription,"power param")
    
    table.insert(performAction,function()
        options={"l","m","a"}
        if behavior.HighPowerAllowed() then table.insert(options,"h") end
        behavior.SetWigglePower("a")
    end)
    table.insert(actionDescription,"power auto")

    table.insert(performAction,function()
        options={"l","m","a"}
        if behavior.HighPowerAllowed() then table.insert(options,"h") end
        behavior.SetWigglePower(options[math.random(#options)])
    end)
    table.insert(actionDescription,"power rand")

    table.insert(performAction,function()
        FindAllMovableSegs()
        cut=NonLockedSegList[math.random(#NonLockedSegList)] --plz only unlocked
        structure.InsertCut(cut)
        behavior.UseCutBands(false)
        seg=NonLockedSegList[math.random(#NonLockedSegList)] --plz not the cut
        options={"h","l","e","a"}
        ss=options[math.random(4)]
        structure.SetSecondaryStructure(seg,ss)
        selection.DeselectAll()
        selection.Select(seg)
        structure.IdealSSSelected() --need cut beforehand in linker
        behavior.UseCutBands(true)
        behavior.SetClashImportance(1)
        structure.WiggleAll(1)
        structure.DeleteCut(cut)
    end)
    table.insert(actionDescription,"random SS, idealSS")

    table.insert(performAction,function() structure.LocalWiggleAll(1) end)
    table.insert(actionDescription,"LwA")

    table.insert(performAction,function() 
        FindAllMovableSegs()
        structure.InsertResidue(NonLockedSegList[math.random(#NonLockedSegList) ])
        FindAllMovableSegs()
        --structure.WiggleAll(1) --debug
    end)
    table.insert(actionDescription,"ins")

    table.insert(performAction,function()
        FindAllMovableSegs() 
        structure.DeleteResidue(NonLockedSegList[math.random(#NonLockedSegList) ])
        FindAllMovableSegs()
        --structure.WiggleAll(1) --debug
    end)
    table.insert(actionDescription,"del. seg, may cause band glitch")

    
    table.insert(performAction,function()
        FindAllMovableSegs() 
        cut=(NonLockedSegList[math.random(#NonLockedSegList) ])
        structure.InsertCut(cut)
        table.remove(NonLockedSegList,cut) --temporary
        if #NonLockedSegList>0 then
            structure.DeleteResidue(NonLockedSegList[math.random(#NonLockedSegList) ]) 
            --hopefully diff.
        end
        --structure.DeleteCut(cut) --can be out of range
        FindAllMovableSegs()
        delCuts() --defined below
    end)
    table.insert(actionDescription,"cut and del. seg") --for 2 sided interface

    table.insert(performAction,function()  
            puzzle.StartOver() FindAllMovableSegs() 
    end)
    table.insert(actionDescription,"reset puzzle. maybe necessary for multi-start puzzles."..         "\n, doesn't exit evolver mode like manual reset does. ") --multi

    table.insert(performAction,function() 
        if multiStart or current.GetEnergyScore()>10000 then 
            puzzle.StartOver() FindAllMovableSegs() 
        end
    end)
    table.insert(actionDescription,"reset puzzle if>10000. for multi-entries contest."..
        "\n, doesn't exit evolver mode like manual reset does. ") --multi

    table.insert(performAction,function() save.Quicksave(3) end)
    table.insert(actionDescription,"save3")

    table.insert(performAction,function() save.Quickload(3) FindAllMovableSegs() end)
    table.insert(actionDescription,"load3")

    
    table.insert(performAction,function() save.Quickload(9) FindAllMovableSegs() end)
    table.insert(actionDescription,"load9")

    table.insert(performAction,function() creditbest.Restore() FindAllMovableSegs() end)
    table.insert(actionDescription,"credit")

    table.insert(performAction,function() absolutebest.Restore() FindAllMovableSegs() end)
    table.insert(actionDescription,"abs best")

    table.insert(performAction,function()
        local ss=math.random(1,structure.GetCount()-3)
        selection.SelectRange(ss,ss+3)
        structure.ShakeSidechainsSelected(1) --All(1)
    end)
    table.insert(actionDescription,"shake(1)")

    function selectWorstSeg()
        
        local lengthOfSegment = math.random(1, 3)
        local segmentInformation = {}
        local workSegmentNumber = math.random(1, 3) --jon asks if Quicksort
        local function sortSegments(segmentInformation)
            for i = 1, #segmentInformation - 1 do
                for j = i + 1, #segmentInformation do
                    if segmentInformation[i][3] > segmentInformation[j][3] then
                        segmentInformation[i], segmentInformation[j] = segmentInformation[j], 
segmentInformation[i]
                    end
                end
            end
            return segmentInformation
        end
        for i = 1, structure.GetCount() - lengthOfSegment + 1 do
            segmentInformation[i] = {i, i + lengthOfSegment - 1, current.GetSegmentEnergyScore(i) }
            for j = 2, lengthOfSegment do
                segmentInformation[i][3] = segmentInformation[i][3]
                     + current.GetSegmentEnergyScore(i + j - 1)
            end
        end
        segmentInformation = sortSegments(segmentInformation)
        selection.DeselectAll()
        selection.SelectRange(math.max(
            segmentInformation[workSegmentNumber][1] - 1, 1), 
            math.min(segmentInformation[workSegmentNumber][2] + 1, structure.GetCount()))
    end
    table.insert(performAction,function()
        selectWorstSeg()
        structure.MutateSidechainsSelected(1)
        selection.DeselectAll()
    end)
    table.insert(actionDescription,"muta worst segment")
    

    table.insert(performAction,function()
        selectWorstSeg()
        structure.WiggleSelected(2,true,true)
        selection.DeselectAll()
    end)
    table.insert(actionDescription,"wig bb sc worst segment")

    table.insert(performAction, function() 
        i=math.random(1,structure.GetCount()) -- do --abridged for debugging 
        structure.SetAminoAcid(i,"e")
    --e for msg 
        --acids arndceqghilkmfpstwyv
    --g for glycine
    --v for valine, suggested by Enzyme for 1947
        --end
    end)
    table.insert(actionDescription,"set amino msg")

    table.insert(performAction, function() 
        bookmark=(bookmark)%structure.GetCount()+1 --7%5=2 ...
        i=bookmark --do --abridged for debugging 
        structure.SetAminoAcid(i,"e")
    --e for msg 
        --acids arndceqghilkmfpstwyv
    --g for glycine
    --v for valine, suggested by Enzyme for 1947
        --end
    end)
    table.insert(actionDescription,"set amino msg bookmark")

    table.insert(performAction, function() 
        i=math.random(1,structure.GetCount()) -- do --abridged for debugging 
        structure.SetAminoAcid(i,"v")
    --e for msg 
        --acids arndceqghilkmfpstwyv
    --g for glycine
    --v for valine, suggested by Enzyme for 1947
        --end
    end)
    table.insert(actionDescription,"set amino val")

    
    table.insert(performAction, function() 
        i=math.random(1,structure.GetCount()) -- do --abridged for debugging 
        structure.SetAminoAcid(i,"g")
        structure.SetSecondaryStructure(i,"l")
    --e for msg 
        --acids arndceqghilkmfpstwyv
    --g for glycine
    --v for valine, suggested by Enzyme for 1947
        --end
    end)
    table.insert(actionDescription,"set amino gly for blueprint")

    table.insert(performAction, function() 
        i=math.random(structure.GetCount())
        aminos="arndceqghilkmfpstwyv"
        choice=math.random(#aminos)
        local char = aminos:sub(choice,choice)
structure.SetAminoAcid(i,char)
    --e for msg 
    
    --g for glycine
    --v for valine, suggested by Enzyme for 1947
       
    end)
    table.insert(actionDescription,"set amino rand")

    
    table.insert(performAction, function() 
        i=math.random(structure.GetCount())
        aminos="arndceqghilkmfpstwyv"
        choice=squeezeToInt(#aminos)
        local char = aminos:sub(choice,choice)
        structure.SetAminoAcid(i,char)
    --e for MSG, q for glutamine
    
    --g for glycine
    --v for valine, suggested by Enzyme for 1947
       
    end)
    table.insert(actionDescription,"set amino param")

    table.insert(performAction, function() 
        aminos="arndceqghilkmfpstwyv"
        i=math.random(#aminos)
        choice=squeezeToInt(structure.GetCount())
        print("choice"..choice)
        local char = aminos:sub(i,i)
        structure.SetAminoAcid(choice,char)
    --e for MSG, q for glutamine
    
    --g for glycine
    --v for valine, suggested by Enzyme for 1947
    end)
    table.insert(actionDescription,"set amino by param")

    --imported from AiLearns22
    table.insert(performAction,function()
        --Bands in Space
        local s1=math.random(structure.GetCount())
        local str=math.random(0.1,1)
            local len=math.random(1,20) --jon: 1-1000
            local theta = math.acos(math.random()) --jon: arccosine
            local phi = 2 * math.pi * math.random()
            local segmentXAxis=0
            local segmentYAxis=0
            while true do --all 3 must be different
                segmentXAxis = math.random(structure.GetCount())--segCnt) --jon
                segmentYAxis = math.random(structure.GetCount())--segCnt) --jon
                if segmentXAxis~=s1 and segmentYAxis~=s1 and segmentXAxis~=segmentYAxis then 
                break end
            end
            --{segmentOrigin, segmentXAxis, segmentYAxis, rho, theta, phi}
            band.Add(s1, segmentXAxis, segmentYAxis, len, theta, phi, str)
           --bands.AddBand(b)
         --puzzle.StartOver()
        
    end)
    table.insert(actionDescription,"band in space")

    table.insert(performAction,function()
        --Bands in Space
        local s1=math.random(structure.GetCount())
        local str=math.random(0.1,1)
            local len=3.5 --angstroms
            local theta = math.acos(math.random()) --jon: arccosine
            local phi = 2 * math.pi * math.random()
            local segmentXAxis=0
            local segmentYAxis=0
            while true do --all 3 must be different
                segmentXAxis = math.random(structure.GetCount())--segCnt) --jon
                segmentYAxis = math.random(structure.GetCount())--segCnt) --jon
                if segmentXAxis~=s1 and segmentYAxis~=s1 and segmentXAxis~=segmentYAxis then 
                break end
            end
            --{segmentOrigin, segmentXAxis, segmentYAxis, rho, theta, phi}
            band.Add(s1, segmentXAxis, segmentYAxis, len, theta, phi, str)
           --bands.AddBand(b)
         --puzzle.StartOver()
        
    end)
    table.insert(actionDescription,"band in space default length")

    act="ends"
--print("act")
    table.insert(actionDescription,act)
    table.insert(performAction, function()   
if band.GetCount() ~= 0 then 
    local band1 = math.random(band.GetCount())
    band.AddToBandEndpoint ( math.random(structure.GetCount()), band1 ) 
end
 end)

    act="for ii = 1, band.GetCount () do band.SetGoalLength ( ii, math.random(0,23.5)) end"
    table.insert(actionDescription,act)
    table.insert(performAction, function() pcall(act) end)


    table.insert(performAction, function() structure.WiggleAll(1) end)
    table.insert(actionDescription, "Wiggle all, "..
        "\n moved from action 0 because Lua starts at 1")

    table.insert(performAction, function() structure.LocalWiggleAll(1) end )
    table.insert(actionDescription, "local wiggle all inspired by BWP")

    table.insert(performAction, function()
        selection.DeselectAll()
        local seg=math.random(structure.GetCount())
        selection.SelectRange(seg,math.max(seg,math.random(structure.GetCount()))) --could use seg+1
        structure.LocalWiggleSelected(1)
        selection.DeselectAll()
    end)
    table.insert(actionDescription, "Local Wiggle Selected inspired by BWP")

    table.insert(performAction, function()
        selection.DeselectAll()
        local seg=math.random(structure.GetCount()-1)
        selection.SelectRange(seg,seg+1)
        structure.LocalWiggleSelected(1)
        selection.DeselectAll()
    end)
    table.insert(actionDescription, "Local Wiggle 2 segment inspired by BWP")

    
    table.insert(performAction, function()
        selection.DeselectAll()
        local seg=math.random(structure.GetCount())
        selection.SelectRange(seg,seg)
        structure.LocalWiggleSelected(1)
        selection.DeselectAll()
    end)
    table.insert(actionDescription, "Local Wiggle 1 segment inspired by Daoism")

    table.insert(
        performAction, 
        function()
            for i=1,30 do
            if structure.InsertResidue(
                math.random(1,structure.GetCount()) --start at 1 for del.
                --I think the wiki is outdated;
                -- insert residue gives an error with index 0 (and count+1) now
            ) then break end
            end
            
        end
        
    )
    table.insert(actionDescription, "Try Insert residue/segment. Only works in Design puzzles.")

    table.insert(
        performAction, function() structure.DeleteCut(math.random(structure.GetCount())) end
    )
    table.insert(actionDescription, "del. cut")

    table.insert(
        performAction,function() structure.InsertCut(math.random(structure.GetCount())) end
    )
    table.insert(actionDescription, "insert cut")

    table.insert(
        performAction, function() 
            structure.DeleteResidue(math.random(2,structure.GetCount()-1)) 
            FindAllMovableSegs()
        end
    )
    table.insert(actionDescription, "delete residue. may cause Foldit crash in combination with GAB")

    table.insert(performAction,function() ui.CenterViewport() end)
    table.insert(actionDescription, "center view") 

    performAction[#performAction+1]=function()
        selection.SelectAll()
        structure.SetSecondaryStructureSelected("a") --auto
        selection.DeselectAll()
    end
    actionDescription[#actionDescription+1]="auto secondary structure."..
     "\n won't affect score until idealSS"

    performAction[#actionDescription+1]=function() --interchangeable here
        --    Set segments secondary structure. Valid values are e=sheet, h=helix, l=loop, m=molecule, a=auto. 
        local struct={[1]="e",[2]="h",[3]="l",[4]="m"}
        structure.SetSecondaryStructure(math.random(1,structure.GetCount()), struct[math.random(1,#struct)])
    end
    actionDescription[#actionDescription+1]="random secondary structure"
    --won't affect score until idealSS

    --end import

    table.insert(performAction,function()
    behavior.SetClashImportance(math.random(0,100)/100)
    end)
    table.insert(actionDescription,"random C.i.") --based on AFK3


    --May need to manually checkmark Enable Recipe Modding in Behavior
    table.insert(performAction,function()
        behavior.SetSidechainHBondImportance(math.random(0,300)/100)
    end)
    table.insert(actionDescription,"side hbond i. 0-3")

    table.insert(performAction,function()
        behavior.SetSidechainHBondImportance(1)
    end)
    table.insert(actionDescription,"side hbond i. 1")


    table.insert(performAction,function()
        behavior.SetPairwiseImportance(math.random(0,300)/100)
    end)
    table.insert(actionDescription,"set pair i. 0-3")

    table.insert(performAction,function()
        behavior.SetPairwiseImportance(1)
    end)
    table.insert(actionDescription,"set pair i. 1")

    
    table.insert(performAction,function()
        behavior.SetPackingImportance(math.random(0,300)/100)
    end)
    table.insert(actionDescription,"set pack i. 0-3")

    table.insert(performAction,function()
        behavior.SetPackingImportance(1)
    end)
    table.insert(actionDescription,"set pack i. 1")


    table.insert(performAction,function()
        behavior.SetHidingImportance(math.random(0,300)/100)
    end)
    table.insert(actionDescription,"set hiding i. 0-3")

    table.insert(performAction,function()
        behavior.SetHidingImportance(1)
    end)
    table.insert(actionDescription,"set hiding i. 1")


    table.insert(performAction,function()
        lastOutputTan=(math.tanh(#structure.GetNote(1))+1)/2
    end)
    table.insert(actionDescription,"getnote")

    
    table.insert(performAction,function()
        lastOutput=#structure.GetNote(2)/10
    end)
    table.insert(actionDescription,"getnote2")

    
    table.insert(performAction,function()
        string=""
        for i=0,math.ceil(lastOutputTan*100) do string=string.."i" end
        structure.SetNote(2,string) --pcall(exit())
    end)
    table.insert(actionDescription,"setnote")

    table.insert(performAction,function()
        behavior.UseCutBands(squeezeToInt(param,2)==2) --not enable
    end)
    table.insert(actionDescription,"usecutbands param")

    --insert actions you want included by default above this line   

    ---------
    --insert new actions here before counting them --jon

    ------old actions don't want to spend debug time below   
    delCuts=function()
        cuts=structure.GetCuts()
        for i=1,#cuts do
            structure.DeleteCut(cuts[i])
        end
    end 
    table.insert(performAction,delCuts)
    table.insert(actionDescription,"del. cuts")

    table.insert(performAction,function()
        absolutebest.Restore()
        delCuts()
        behavior.SetFiltersDisabled(false)
    end)
    table.insert(actionDescription,"get credit for abs")

    table.insert(performAction,function()
        behavior.SetClashImportance(1) --needed?
        behavior.UseCutBands(true) --not enable
        structure.WiggleAll(1)
        cuts=structure.GetCuts()
        for i=1,#cuts do
            --structure.WiggleAll(1)
            structure.DeleteCut(cuts[i])
        end
    end)
    table.insert(actionDescription,"del. cuts2")
    -------------------
    
    NumberOfActions=#performAction+1 --jon

    save.Quicksave(9)
    if not testXor and debugActions then
    for iParam=1,1 do
    param=math.random(-100,100)/10
    print("debug actions, param "..param)
    for i=1,#performAction-1 do --i=1 to skip mutate
        --save.Quickload(9) --komi may help
        starttime=os.clock()
        startscore=current.GetEnergyScore()
        print("debug action "..i..", "..actionDescription[i])
        performAction[i]()
        print("yield per sec "..(current.GetEnergyScore()-startscore)/(os.clock()-starttime+1))
    end
    --structure.WiggleAll(4) --lag for debug
    end 
    end
end
initializeActions()
----------------
pokemons={}
--perhaps bias (threshold) is unfortunately named that way
--https://stackoverflow.com/questions/2480650/what-is-the-role-of-the-bias-in-neural-networks
--it means shifting a value to one side in algebra
--originally from French:
--https://www.etymonline.com/word/bias

numInputs=10
--must be exactly the same as number of inputs below

for i=1,4 do --*2 do --ought to be derivative of (numDendrites+1 [for bias])*numNeurons
    local pokemon={layers={},score=0}  --"ogres are made of layers"

    function newNeuron(--[[optional]]numDendrites)
        newDendrites={}
        numDendrites=numDendrites or numInputs+1 --at least size of biggest layer
        --https://forum.defold.com/t/lua-defining-a-function-with-optional-parameters-solved/3340
        
        for iDendrite=1,numDendrites do --maybe extra for now?
            table.insert(newDendrites,math.random(-10,10)/10)
        end
        local neuron={
            dendrites=newDendrites
            --weights (dentrites getting inputs from prev. neurons)
            -- are factors to have a weighted importance
            --for instance, if you're guessing the total weight of 2 elephants and 2 insects
            --the number of elephants will be far more important
            ,
            bias=math.random(-10,10)/10--can exceed
            --bias in this case is like a T/F threshold. <50% on exam = F, you don't pass at all. 
            --some professors curve grades based on overall class grades
            ,
            axon={0,1} --called neuron Value by techies. means neuron output
        }
        --print("neuron axon "..neuron.axon)
        return neuron
    end

    layer={} --input layer, true, false
    for i=1,numInputs do 
        table.insert(layer,{axon={0,1}})
    end
    print("input neuron axon "..layer[1].axon[1])
    table.insert(pokemon.layers,layer)

    for iHiddenLayer=1,hiddenLayers do
    layer={} --new layer
    --a Hidden Layer
    --for XOR problem, set #neurons to 2
    -- https://miro.medium.com/max/1346/1*bRoaMcjiD4yRjgfr2QGZcA.jpeg
    for i=1,#pokemon.layers[1]+1 do --3 do --2 do
        local neuron=newNeuron()
        --print("my neuron "..neuron)
        table.insert(layer,neuron)
    end

    print("done inserting neurons into hidden layer")

    table.insert(pokemon.layers,layer)
    end

    layer={} --new layer
    --output layer
    for i=1,2 do
        local neuron=newNeuron()
        table.insert(layer,neuron)
    end
    table.insert(pokemon.layers,layer)

    table.insert(pokemons,pokemon)
    
end
gen=0
lastOutputTan=.2 --forgot this

lastOutput=0
weakest=1

print("initializing velocity")
    mutantLayer=math.random(#pokemons[1].layers-1)+1
    mutantNeuron=math.random(#pokemons[1].layers[mutantLayer])
    mutantDendrite=math.random(#pokemons[1].layers[mutantLayer][mutantNeuron].dendrites)
    dendriteMutation=math.random(-50,50)/100 --Mozi's precision philosophy
    biasMutation=math.random(-50,50)/100 --Mozi's precision philosophy
while true do
    --88% cpu no recipe running
    --100% cpu just checking hour
    --140% cpu active Neural Net
    print("sleeping")
    repeat
        hour=(os.date("%H"))
        --hour="00"
        --hour="23"
        day=(tostring(11)<hour and hour<tostring(19))
    until true or not day
gen=gen+1
print("new gen "..gen)
path=1
for i=1,#pokemons do
    starttime=os.clock()
    --creditbest.Restore()
    --band.DeleteAll()
    --behavior.SetFiltersDisabled(false)
    startscore=current.GetEnergyScore()
    if not current.AreConditionsMet then 
        startscore=worstPossible --0 
    end
    
    --puzzle.StartOver()
    --puzzle.StartOver() --if multistart
    diversity=0
    pokemons[i].score=worstPossible --0 --for xor. for foldit, set to worst, get highscore?
    if testXor then pokemons[i].score=0 end
    attacks=4 --try even number for xor/not
    iAttack=1
    while iAttack<=attacks do --9 do
        
    path=1--math.random(1,2) --set for input
    selection.DeselectAll()
    --selection.SelectAll()
    selection.SelectRange(math.random(structure.GetCount()-3),math.random(structure.GetCount())) 
    --hopefully ends mutable if design but sadly not linker

    --randomize inputs to test XOR
    pokemons[i].layers[1][1].axon[path]=(iAttack-1)%2 --math.random(0,1)
    pokemons[i].layers[1][2].axon[path]=math.floor((iAttack-1)/2) --math.random(0,1)

    pokemons[i].layers[1][3].axon[path]=behavior.GetClashImportance() --math.random(0,1)
    pokemons[i].layers[1][4].axon[path]=0--param --lastOutputTan
    pokemons[i].layers[1][5].axon[path]=iAttack --algoStep)
    pokemons[i].layers[1][6].axon[path]=0 --
    if #structure.GetCuts()>0 then
        pokemons[i].layers[1][6].axon[path]=1
    end
    pokemons[i].layers[1][7].axon[path]=(scoreboard.GetRank()) --shame,accurate endgame detection
    pokemons[i].layers[1][8].axon[path]=0 --overkill updating
    if HASMUTABLES then
        pokemons[i].layers[1][8].axon[path]=1
    end
    pokemons[i].layers[1][9].axon[path]=((1+current.GetEnergyScore())/2) --
    pokemons[i].layers[1][10].axon[path]=0--lastOutput
    if testXor then 
        for iAxon=3,#pokemons[i].layers[1] do
            pokemons[i].layers[1][iAxon].axon[path]=0 
        end
    end

    --we don't really understand bifurcated axons, so we generally don't simulate them
    path=math.random(1,2) --randomize for road less traveled if even amount of neurons
    if humble then path=1 end

    --skip activation of first layer (input layer)
    for iLayer=2,#pokemons[i].layers do 
        print("gen "..gen..", pokemon #"..i..", layer "..iLayer..",iA"..iAttack)
        --print("layers is "..pokemons[i].layers)
        print("layer length "..#pokemons[i].layers[iLayer])
        for iNeuron=1,#pokemons[i].layers[iLayer] do
            --"inputs go in times weights, summate, add a bias activate" --s. raval
            
            sum=0
            for iPrevNeuron=1,#pokemons[i].layers[iLayer-1] do --previous neurons
                if iLayer>2 and not humble then 
                    path=3-path --road less traveled 
                end
                local input=pokemons[i].layers[iLayer-1][iPrevNeuron].axon[path]
                local dendrite=pokemons[i].layers[iLayer][iNeuron].dendrites[iPrevNeuron]
                --print("input "..input..", neuron weight "..dendrite)
                sum=sum+input*dendrite
                --input times weight of connection
            end
            print("sum "..sum)
            --"inputs go in times weights, get sums" -s. raval

            print("neuron bias "..pokemons[i].layers[iLayer][iNeuron].bias)
            
            local axon={0,0}
            axon[path]=sum+pokemons[i].layers[iLayer][iNeuron].bias
            --most people add bias instead of subtract bias but maybe it's easier visualizing 
            -- subtracting bias, so you can view the bias as a threshold, then you can write
            -- if axon<=bias instead of axon<=.5
            -- or something like axon<=threshold

            --"pass that **** [stuff] to my sigmoid function" -s. raval

            --axon=math.tanh(axon)
            
            -- [[
            if iLayer<#pokemons[i].layers then --allow output to be negative
            --ReLU activation function, not really a sigmoid function like tangent
            if axon[path] <=0 then --.5 then
                --axon[path]=0 --ReLU, rectified linear unit. dying ReLU
                --  graph: __/
                aloha=0.5 --alpha
                axon[path]=aloha*axon[path] --leaky relu

            end
            end
            --]]--

            --MIT OpenCourseWare video of training a NN to output the same as the input
            -- on MIT's site
            --bifurcated axons sometimes
            --more like chopsticks/kuazi than a fork, because not as branched as dentritic tree
            pokemons[i].layers[iLayer][iNeuron].axon[path]=axon[path]
            pokemons[i].layers[iLayer][iNeuron].axon[1-path]=0
        end
    end

    local brainOutput=pokemons[i].layers[#pokemons[i].layers][1].axon[path]
    lastOutput=brainOutput --not sure if nn can handle inputs>1,<0,<-1
    print("gen"..gen..",brainOutput: \n"..brainOutput..",\n param "..param)
    newTan=(math.tanh(brainOutput)+1)/2  --0 to 1
    diversity=diversity+math.abs(lastOutputTan-newTan)
    --print(#performAction)
    lastOutputTan=newTan
    print("output tan "..lastOutputTan)
    choice=math.floor(brainOutput%#performAction)+1 --pemdas
    --big out of bounds fix... and now maybe fixed for too negative for float too
    --skip action 0 for now (mutate all), simplify math

    --originally
    --choice=math.floor((brainOutput*(1+#performAction))%(1+#performAction))--+.49)*lastOutputTan)
    --causes crash when brainOuput is 2.58e+306, guessing because it wraps to negative

    print("choice"..choice..", #actionDescription"..#actionDescription)

    
    if chooseAminos and not testXor then
        
        param=pokemons[i].layers[#pokemons[i].layers][2].axon[path]

        --inspired by PUBG
        if param>10000 or param<0 then 
            startscore = startscore+math.abs(param)
            param=0 
        end
        if brainOutput>10000 or brainOutput<0 then 
            startscore=startscore+math.abs(brainOutput)
            brainOutput=0 
        end

        aminos="arndceqghilkmfpstwyv"
        local myAmino=math.floor(brainOutput%#aminos)+1
        print("my amino is "..myAmino)
        seg=math.floor(param%structure.GetCount())+1
        local char = aminos:sub(myAmino,myAmino)
        print("char is "..char)
        structure.SetAminoAcid(seg,char)
        print("set amino succ")
        --e for MSG, q for glutamine
    
        --g for glycine
        --v for valine, suggested by Enzyme for 1947
        --attacks=attacks+1
    else
        print(actionDescription[choice])
        if (not testXor) and 0<choice and choice<=#actionDescription then 
        
            param=pokemons[i].layers[#pokemons[i].layers][2].axon[path]
            performAction[choice]() --paused for xor debug
        
            if param==1 and math.random(90)>1 --arcade coins continue? 
                then attacks=attacks+1 
            end --jiexu
        end
    end
    
    komi=i--*6.5 --komidashi for playing 2nd

    improv=(current.GetEnergyScore()-startscore)/(os.clock()-starttime+1)--+diversity+komi
    if not current.AreConditionsMet() or #structure.GetCuts()>0 or behavior.GetFiltersDisabled() then
        improv=improv*0.0 --7 --0

        --*0.7
        --shifu: 
         --layer , neuron: dendrites: -1.642, 1.74, -2.005, -4.18, , bias: 3.048, 
        --neuron: dendrites: -5.839, 3.608, 1.621, -2.22, , bias: 3.683
        --layer , neuron: dendrites: -0.023000000000014, 0.67800000000008, , bias: -1.7139999999999
    end
         
    function orFunc(a,b) return math.min(1,a+b) end
    function andFunc(a,b) return a*b end
    function notFunc(a) return 1-a end

    -- https://miro.medium.com/max/1346/1*bRoaMcjiD4yRjgfr2QGZcA.jpeg
    xorFunc= function(in1,in2) 
        return andFunc(
        notFunc(andFunc(in1, in2)),
        orFunc(in1, in2)
        )
    end

    --real vs predict
    path=1 --for xor
    print("xor func "..xorFunc(0,0)..xorFunc(0,1)..xorFunc(1,0)..xorFunc(1,1))
    print("pokemons[i].layers[1][1].axon[path]"..pokemons[i].layers[1][1].axon[path])
    print("pokemons[i].layers[1][2].axon[path]"..pokemons[i].layers[1][2].axon[path])
    xorReal=xorFunc(pokemons[i].layers[1][1].axon[path],pokemons[i].layers[1][2].axon[path])
    print("xorReal success")
    xorPred=pokemons[i].layers[#pokemons[i].layers][1].axon[path]
    error=xorPred-xorReal
    
    realNot=notFunc(pokemons[i].layers[1][2].axon[path])
    predNot=pokemons[i].layers[#pokemons[i].layers][2].axon[path]
    --print(""..notFunc(0)..notFunc(1))
    --error=predNot-realNot    

        
    if testXor then
        pokemons[i].score=pokemons[i].score -math.pow(error,2)/2/attacks
         behavior.SetClashImportance(
        --math.min(1,math.max(0,(pokemons[i].score+10)/10)) --for visual
        --squeezeFloat(0,1)
            eulerAct(pokemons[i].score+1)
        )
   
    else
            --alpha for score rectifying, eg. 0.1
        pokemons[i].score=
        
            math.max(pokemons[i].score,math.max(0.0*improv,improv))--+komi -- 
    end
    print("xorPred "..xorPred)
    print(",xorReal "..xorReal)
    print("predNot "..predNot)
    print("realNot "..realNot)

    --structure.WiggleAll(1) --debug
        iAttack=iAttack+1
    end --iAttack

    print("score "..pokemons[i].score..", 0 is best")
    --structure.WiggleAll(1) --lag for debug
end
strongest=1
for i=2,#pokemons do
    if pokemons[i].score>pokemons[strongest].score then strongest=i end
    if pokemons[i].score==pokemons[strongest].score and math.random(2)>1 then 
        strongest=i end
end
print("strongest "..strongest..", score "..pokemons[strongest].score)
if weakest==strongest then 
    print("gen"..gen..",carp evolved to dragon success") 
    dendriteMutation=dendriteMutation*1.2
    biasMutation=biasMutation*1.2
    if testXor then
    timer=os.clock()+5
    while (os.clock()<timer) do
        structure.WiggleAll(1) --debug
    end
    end
else
    if true or math.random(2)>1 then --chance of tripping, chance of carrying momentum downhill on
    mutantDendrite=mutantDendrite+1
--math.random(#pokemons[1].layers[mutantLayer][mutantNeuron].dendrites)
    print("if you give a mouse a cookie")
    if mutantDendrite>#pokemons[1].layers[mutantLayer][mutantNeuron].dendrites then
        print("it's gonna wanna layer")
        mutantNeuron=mutantNeuron+1 --math.random(#pokemons[1].layers[mutantLayer])
        mutantDendrite=0
        if mutantNeuron>#pokemons[1].layers[mutantLayer] then
            mutantNeuron=1
            mutantLayer=mutantLayer+1
            print("if you give a mouse a wafer")
            if mutantLayer>#pokemons[1].layers then 
                mutantLayer=2 --back to first hidden layer
                yellowWood=-yellowWood --Road Less Traveled
            end
            
        end
    end
            
    dendriteMutation=0
    biasMutation=0
    if mutantDendrite>=1 then
        dendriteMutation=yellowWood*math.random(50)/100 --Mozi's precision philosophy
    else
        biasMutation=yellowWood*math.random(50)/100 --Mozi's precision philosophy
    end
    end
end

weakest=1
for i=2,#pokemons do
    if pokemons[i].score<=pokemons[weakest].score then 
        weakest=i
    end
end
runnerup=weakest --2nd best
secondWeakest=strongest
for i=1,#pokemons do
    if i~=strongest and pokemons[i].score>=runnerup then
        runnerup=i
    end
    if i~=weakest and pokemons[i].score<=secondWeakest then
        secondWeakest=i
    end
end
print("weakest "..weakest..", evolving weakest carp to dragon")
for w=1,1 do --2 do --only 1 for gradient simplicity for now
i=weakest
if w>1 then i=secondWeakest end
shifu=strongest --in Kung Fu Panda, Shifu is the 2nd wisest
--shifu meant tutor; now it means driver/chef
    shifustring="shifu: "
--skip copying of first layer (input layer)
    for iLayer=2,#pokemons[i].layers do
        print("pokemon #"..i..", layer "..iLayer)
        --print("layers is "..pokemons[i].layers)
        --print("layer length "..#pokemons[i].layers[iLayer])
        shifustring=shifustring.."\n layer "
    
        for iNeuron=1,#pokemons[i].layers[iLayer] do
            --shifu can be female, shimu means tutor's wife but gender is a hassle to code
            --plus most CandyCrush pros are female
            --shizhang can mean tutor's husband
            local shifuNeuron=pokemons[shifu].layers[iLayer][iNeuron]
            local shimuNeuron=pokemons[runnerup].layers[iLayer][iNeuron]
            noob=pokemons[i].layers[iLayer][iNeuron]
            shifustring=shifustring..", neuron: dendrites: "
            --we could do backpropagation, involves finding the policy gradient/slope
    
            for iPrevNeuron=1,#pokemons[i].layers[iLayer-1] do --previous neurons
                local shifuDendrite=shifuNeuron.dendrites[iPrevNeuron]
                local shimuDendrite=shimuNeuron.dendrites[iPrevNeuron]
                print("shifuDendrite weight "..shifuDendrite)
                shifustring=shifustring..shifuDendrite..", "
                if math.random(2)>1 then
                    allele=shifuDendrite
                else
                    allele=shimuDendrite
                end
                
                --gradient descent
                mutation=0
                if iLayer==mutantLayer and iNeuron==mutantNeuron and iPrevNeuron==mutantDendrite then
                    mutation=dendriteMutation end
                pokemons[i].layers[iLayer][iNeuron].dendrites[iPrevNeuron]=
                    --tryhard disciple method
                    --noob.dendrites[iPrevNeuron]+(shifuDendrite-noob.dendrites[iPrevNeuron])*2

                    --pro vs pro
                    --shimuNeuron.dendrites[iPrevNeuron]
                      --  +(shifuDendrite-shimuNeuron.dendrites[iPrevNeuron])*2

                    --valentine method
                    --(shifuDendrite+shimuDendrite)/2

                    --other half method
                    --allele

                    --unpressured aphid method
                    shifuDendrite

                    +mutation
                    
            end

            print("shifu's orig bias "..shifuNeuron.bias) --shifu not biased against MSG
            
            shifustring=shifustring..", bias: "..shifuNeuron.bias
            local allele
            if math.random(2)>1 then
                allele=shifuNeuron.bias
            else
                allele=shimuNeuron.bias
            end
            mutation=0
            if iLayer==mutantLayer and iNeuron==mutantNeuron then
                mutation=biasMutation end
            pokemons[i].layers[iLayer][iNeuron].bias=
                --tryhard disciple method, kinda velocity gradient descent
                --noob.bias+(shifuNeuron.bias-noob.bias)*2
                --Tutor says to chop a log, chop 2 logs
                
                --tryhard pro method, note that I don't mean tryhard in the derogutory sense
                -- but rather that imitation is a fast (albeit not as great as AlphaZero) method
                --After all, I guess it's why stories unrealistically have happy endings
                --(shimuNeuron.bias+(shifuNeuron.bias-shimuNeuron.bias)*2)
                --Eg. What can Starcraft pro Scarlett do to play like #1 Serral but better?
                --Double Evo chamber or more creep tumors?
                --This method finally got the a.i. to set clash 1

                --valentine method, but kinda a social bubble stagnation
                --"2 roads diverged in a yellow wood and the a.i. car barged down the middle" -GeoHot
                --(shifuNeuron.bias+shimuNeuron.bias)/2
                
                --half of genes method
                --allele

                --aphid method
                shifuNeuron.bias

                +mutation
                
        end
    end
    --backpropagation recurses from the output to the input neurons,
    -- updating them according to how responsible they are for the error
    -- siraj r.

    shifustring=shifustring.."\n " --easier selecting text
    structure.SetNote(1,shifustring) --middle for 2 sided interface? --1
    --max locked seg for 2 sided? stays on even if a res is del.

    --2 sided
    --shifu: 
     --layer , neuron: dendrites: 0.17, 0.578, 0.282, 0.374, , bias: 0.068, 
    --neuron: dendrites: 0.711, 0.286, 0.947, 0.777, , bias: -0.51
     --layer , neuron: dendrites: -0.91, 0.214, , bias: 0.376


    --1951 cov linker
    --shifu: 
     --layer , neuron: dendrites: 0.812, 0.855, 0.526, , bias: -0.253, neuron: dendrites: -0.325, -0.97, 0.955, , bias: 0.627
     --layer , neuron: dendrites: 0.694, 0.905, , bias: -0.826

end
--print("end carp to dragon?")
--sometimes get bias of -0 negative zero. but -0==0 in lua.
end
--------