🖥️ FINISHING MY FIRST MACHINE LEARNING GAME! (3/4)

[@AJGaeming]

9:10 Did the audio desync in the editing phase, since it is the same audioclip from the beginning of the video? :s Then it cuts to music mid-word at 11:48

[@Jabrils]

20 days after upload someone claimed this video for music, KZbin asked me if I wanted to let them remove it, I said yes so that I can keep the video monetized, but after they edited it, that is what it did. Now I am stuck without being able to revert back to the original with no KZbin contacts to get help from. I've added an alternate upload here's the link: kzbin.info/www/bejne/j2mTp32uod5lfdU. I'm sorry. - Jabril

[@AJGaeming]

Ah, gotcha! It's ok, I was just confused and thought I was having a massive deja vu :D Great video nonetheless!

[@ReignSurvives]

This has been my fear on my own videos as well. It's pretty strange how they chose to edit it, but good job making a fix. It might pay to hide this version and then make sure the other one is in the playlist or is linked to from episode 2/4. Oh the headaches youTube creates. I'm loving the content. Please keep making these.

[@Jhaskydding]

same thought

[@demohub]

Please add these slices to the Threelly SmartView Chrome extension. chrome.google.com/webstore/detail/threelly-smartview-for-yo/dfohlnjmjiipcppekkbhbabjbnikkibo/

[@jimmylewis7896]

Jabril your content is extremely high quality. You make computer science interesting to a guy like me, who doesn't know anything about it

[@Jabrils]

hey! that's an honor. thanks for sharing that Jimmy. - Jabril

[@iVeggie]

@@IbrahimSharara1 LOL

[@scrapeape]

Far and away the best , most in-depth (and understandable to a layman enthusiast) insight into an ML programmer's thought process that I've seen on KZbin. Thanks!

[@lukealdrich]

you should make another ai to automatically build harder courses any time forest completes them.

[@AceofSpadesss]

I read somewhere in bible that this is how the end of the world begins

[@dankmemes7658]

@@AceofSpadesss lmao yeah sure. I'd love to see that verse

[@preppertechnicianee6013]

@@dankmemes7658 the lord sayeth I created man in my image with a biological neural networks But he ate of the tree of knowlegde Now hes gonna fuck it all up by programing ....... Me

[@scottmaxwell1927]

@@dankmemes7658 lol spat out my coffee reading your reply :)

[@tertialmule3309]

damn I was going to comment that xD

[@RedEyedJedi]

I absolutely loved this series. Thanks for making it buddy. I'm glad I'm not the only one who finds this topic fascinating.

[@imnotselma3305]

Watching Forrest trying to run the different courses has been surprisingly mesmerizing. I'm gonna miss the little guy

[@Jabrils]

😢😢😢 i promise im chopping onions. - Jabril

[@creationsmaxo]

Really cool short series here. What's quite awesome is thinking about about additional methods/ways that could have been implemented to complete and make it even more efficient. For example, adding normals analysis onto the raycasts that determines the scores on the 5 directions. Getting the normals returns the directions from which the 3D faces (walls) are facing. That's, after all, the principle behind how we're able to follow a track ourselves : We see the distances and directions from restriction we're facing. Even someone who's blind uses the same principle by using his ears to analyse the direction of hard surfaces and the echos. (sounds doesn't return the same if it bounce at 90 degrees off a surface or less.)

[@I3lazeKing]

Cheers for making this - it was a good watch! I've played with machine learning and haven't had the chance to delve into deep learning yet but this series helped visualise it better :) Keep up the good work!

[@ti83magic]

Your Chanel got randonly suggested to me by youtube, and I was glad it was. Very entertaining to watch, and you did a vey good job at explaining the concepts behind the algorithms. A Well-deserve subscription!

[@BrunoGocan]

Your level of mental energy is out of this world!! I have no words, I'm a fan now!!

[@hendrixkid2362]

I think the random slice works best because it is not limited by the past results, and potentially adds new arguments that could totally work! I loved this video, you have gotten me excited about machine learning! I'm going to talk to my academic adviser about switching my major!

[@jonsirmans1557]

At first I thought your videos were silly. I am a fan now. Subbed 100%. I hope you grow because you're super entertaining. Good luck!

[@vesperannstas]

Wow, that race really had me on the edge of my seat!

[@stupidrainbo]

I'm writing this out before watching the post-mortem, so forgive me if I'm saying things you already went over... I think there are a couple reasons the results weren't as good as they could have been. First, it seems like the turn radius isn't sharp enough to turn around. So when Forrest runs up to a confusing intersection he hasn't seen before, there is no way to guess whether it goes left or right. He just has to pick one, and if it's a dead-end, that simulation loses because he won't be able to turn around. However, that simulation didn't lose because it's genetically unfit -- it only lost because it was genetically predisposed to guess left at the confusing intersection. This means the automatic assumption by the computer is that he was genetically unfit for any race in general, but really he was just guessing that the intersection turned left. In fact, that could have been the best Gump yet, but lost because he rolled the dice wrong. Second, piggybacking off that first point, the A.I. may just be training by memorizing the track. Who knows if it's actually learning how to navigate new areas? All it knows is the specific track. Every time you train it on a new track, it is going to fail in the beginning and then have to readjust to the new track as it memorizes the new one and forgets the old. Ideally, I think, the best way to train it would be to 1) Give it a sharp turn radius, and 2) Randomly generate a new track for every generation. That way, it has to rely on its ability to adapt 100% and not at all on its ability to remember tracks. Of course, then it would be hard to reward it properly for advancing in the track so 3) You'd have to give it an incentive for getting closer to the end goal... maybe draw an A* path to the end and make the distance inverse to the score.

[@patriciafergus9929]

Way to call the election.

[@ianborukho]

Just realized I couldn't see the comment option because window was too small XD. These vidz are so dope. Clarity, content, reasoning, animation, editing, acting, music choice, dragonball reference, and overall entertainment value are 100. Thanks.

[@bringamosa8480]

Dude i am NOT able to do anything with programming, somehow i came across your channel, saw this series, and must say i really really like your way of doing this. Informative and funny. keep it up!

[@mr.peanutbutter303]

Dude honestly, if you don't continue to make educational videos, it will be a crime. You are amazing. Thank you

[@Trudas00]

lol it's actually strange to see him at the end of the video actually talking haha ... good job man, nice videos

[@dangel9293]

Just saw this series of videos on Machine learning. Very interesting and i'm enjoying them quite a bit. Just wanted to note that I think the Random % learning selection is actually the best way to mimic life.

[@yusefalimam130]

BY FAR THE BEST IVE FOUND FOR GETTING INTO IT THE BEST

[@geordonworley5618]

When it comes to crossover, if crossover is performed on the whole population it will effectively make them all "the same species". Every time two networks with different strategies are mated, it is almost certainly going to result in failure (nonviable offspring), so the resulting "species" all work the same so their pieces are interchangeable with slight variation. That is highly consistent with what you are seeing. I see a lot more success in simulations I've ran when I allow speciation to happen if I am using crossover or if I allow species to have "crossover points" that allow them to crossover at some points if it matches with the one being mated with. However, in your case it seems that since there is just one solution to this problem that only one species was required anyways, so it made more sense to just make the whole population quickly be that.

[@jay-wz4wc]

Found this video (and your channel) from a tweet of another programming youtuber and I absolutely love it keep up the awesome work man!

[@rman1572]

I greatly underestimated how entertaining this series would be, awesome! subscribed and know to check all your other content.

[@bensaxon9206]

I love your videos. First off, you get to the facts in a really fun an interesting way. Your have such a great character (no pun intended, Forrest), and overall your a really cool guy. I give you my subscribe.

[@Snakebite420]

sooooo good, I love how you set up these videos as entertaining and easy to follow along. heck honestly im not even doing this, but now i kinda want to. Im still going to watch the whole series today and i've subscribed. Thank you for peaking my interest again, this coming from a college student going for a Computer Science Degree.

[@raul825able]

Hey Jabrils, Thank you for this great content. This is my fav channel now on :)

[@guslarscheid3606]

Episode 2: Don't talk to me about politics. Episode 3: Brings up politics.

[@M3A7]

IKR.... Effff

[@johnplus7069]

Gus Larscheid boohoo

[@meowinferalas]

Yo trump is bs not politics

[@M3A7]

He's the president, so you're a dumb fuck. Grats!

[@Nae_Ayy]

Either way he's still an incompetent piece of shit.

[@corvusclones]

one likely reason for natures adaptation of the slice crossover is that nucleic acids aren't single input values, they come in triplets called codones that correspond with amino acids for protein building. Because of the length of polypeptides, switching a single nucleic acid could have monumental consequences, like accidentally making tons of stop codones which says to stop the polypeptide construction after tens of amino acids for proteins that would normally be hundreds or thousands long

[@corvusclones]

also, an easy way to prevent loss of training progress is to make a clone of the best of each generation in the next generation, that way, if some special quality that made it particularly effective would otherwise be bred out in the next generation, it's preserved

[@bulka2914]

this man has the hand of god, I'm now blessed with machine learning.

[@AnuragSyal]

This is genius work. Pure knowledge. You should be writing a paper on this for a game-dev conference!!!

[@grindersgaming9447]

Man I absolutely love this channel to bits!

[@naughtnoa6185]

Definitely my new favorite channel. You're hilarious.

[@Wownerd1265]

5:00 keep in mind also, nature has not necessarily produced the best possible system for evolving. It has created a system that has survived for a long time and (subjectively speaking) done pretty well for itself, but it is probably not the best method possible.

[@beachboardfan9544]

Soooo could this be applied to my lawn mower so it'll cut my grass for me?!

[@Jabrils]

yes it can. infact, if you get started on this i am sure millions of dollars are waiting for you haha. delete this post quick! - Jabril

[@skolex3121]

Sure, just remember to keep the door closed so it won't mow the ground floor of your house as well.

[@eddyleon1665]

skolex but then it might go and mow everyone else's lawn too (for free). we can't have that

[@beachboardfan9544]

Ehh just put a tip jar on it...

[@beachboardfan9544]

hahaha I'm slightly concerned for the world that it took someone three weeks to think of the dog!

This guys is a Damn Genius!

[@todea]

I love this, thank you for sharing your knowledge and making such awsome content. This really motivates me to learn about machine learning.

[@ugosouza2127]

I believe that even if I were not already interested on the subject, I would be after watching you dude. Loved it!

[@silentgrove7670]

I immediately thought of a GAN. Another network builds harder and harder courses for Forrest to try. That would be a super cool project. Lately I wonder if my life is like that. Reality is presenting me more complex problems to solve. Thank you universe for keeping me challenged !!!

[@SuperbossGames]

That ff7 music though

[@Jabrils]

had to do it :) - Jabril

[@fakhruz3105]

@@Jabrils that made me cry a bit

[@anthonyp9124]

Your style is so unique! I love it

[@lanha3385]

you really inspire people to learn machine learning

[@Rorypandanda]

Hey Jabril! Awesome videos man. Keep up the good work!

[@Majkeel]

WTF 1500?! Your channel is highly underrated. You get all of my views. Thank you for vids, I love 'em

[@karan4624]

I love his way of speaking😍

[@insideTheMirror_]

Was I the only one who got too excited to see him talking straight to the camera?

[@gshap1258]

This one project has so many different types of jobs in it. Data science, AI, Machine learning...like wtf. How smart can you be

[@numatechprototypes222]

See how you get around the whole decision making for random or precise functions is using your random function as a optimization side loader that runs a real-time micro SIM of the scenario going on using its top performing data and communicating it back to the precise function that operates the main character or scenario. And what you'll see in the graph is an exponential rise in performance basically you're giving me AI the ability to self-simulate and communicate the best data to the motor control function

[@peterpan872]

"Take a second to let that register...." 😂😂😂 I’m weak

[@angelogrant6749]

I love the progress that you made with this project. I have a few helpful suggestions that may work to help you. First, I think you would need to train Forest to run both ways. Second, you will have to store the information of what forest learnt (in a database of some type) from previous course that he ran so he can apply the knowledge that he learnt for his new course that he's about to run. Hope you see my logic.

[@TheLunaLockhart]

"the data says A is the best way forward, but let's also try B" Spoken like a true scientist! ALWAYS seek to disprove things you observe, especially when you are looking at an emergent field

[@NaudVanDalen]

The link to part 2 in the description actually links to part 1. Also it would be nice if the description had the link to the next video. Or make video end cards, because KZbin doesn't always recommend the next video and people are not always watching the playlist. That increases views and watch time and retention.

[@mcasualjacques]

dont know if you mentioned it but an age old strategy to visit a labyrinth ( which i used in Wolfenstein ) is to go forth, keeping the right ( or left ) hand on the wall

[@Try_Hard_Dad]

Love your stuff bruv!

[@AtticusDragon]

Nature easily marries with your data suggesting processes with randomized values are the most efficient methods of learning. Consider that genetic mutation is practically random - it's not that a species will intelligently adapt to environmental factors, but rather those which develop advantageous traits due to random genetic mutations are more likely to survive and thus pass on the new characteristics. Conceptually it's near identical to your randomized processes recieving priority values, favoring efficient processes and passing the data to the next function. It is interesting all the same that these processes would be so much more efficient when compared to processes that had intelligent guidance.

[@EDToasty]

Also whos to say that slicing this away is used in nature? It doesn't actually matter which weights / biases are selected, because the weights and biases can be arranged any way you really want to.

[@wr41thx29]

Interesting. The random approach tests more possibilities because it sometimes just doesn't follow a particular form of logic...well besides the logic that dictates the generation of said randomness. That makes me wonder if different seeding methods for random number generation would have an impact.

[@Arelias95]

When you do crossover, you should have a random chance of mutating the output, as this does happen in nature, maybe with that it could beat the random splicing approach :)

[@BillionPlusOne]

I am super enjoying this. Awesome stuff! (I did a BSc in artificial intelligence 25 years ago - I wish I had cool tools like these to play with!)

[@BiggerMikey]

Love these videos, man. Your explanations and sense of humor make this kind of information fun and easily digestible. What aren't fun and easily digestible are Whoppers. Dude. Whoppers are trash. Keep it up!

[@danieldbird]

How did you make the hex courses? Is it in unity? Or something else?

[@zeroxenator]

Hey, Whats your machine specs? Looks like you are using Dell XPS 15" inch for the simulations?

[@ThatGuy53297]

Could you change the probability method halfway through? If you start with the random one and get a long high then a big drop, could you take the series of high ranked numbers and use the slice on them?

[@nickspringham9989]

What software do you use to animate your videos? p.s. this is a sick video series!

[@Sudonym]

I thought I was going Insane for a Minute on 9:10. Guessed I had another Tab with that Video open, luckily I saw the pinned Comment ^^ Anyways, Array Index starts at 1. Remeber folks.

[@demohub]

Please add these slices to the Threelly SmartView Chrome extension. chrome.google.com/webstore/detail/threelly-smartview-for-yo/dfohlnjmjiipcppekkbhbabjbnikkibo/

[@DaveDDD]

In nature, random crossovers are limited to mutations; they're rare and are often not beneficial, and too much of a deviation means the organism won't be able to survive or even live in the first place. If a species has too many offspring that are unable to survive and propagate their genes, it will die out entirely. A computer program doesn't have the problem of needing to spawn children that are able to make it to breeding age, so it's a better tactic to "go for broke" with random crossovers to beat out the local minimum (as you mentioned).

[@AgglomeratiProduzioni]

6:53 I removed my upvote just to give it again.

[@ereyes95]

I am jealous of your brain. Well done!

[@KevonLindenberg]

6:44 dat explanation of random crossover function, lolol.

[@rikvandenreijen]

Just came across your channel. Have to say that you have an interesting style of editing. Keep up the great work!

[@tbhalo]

Could you give some advice about how you got from, wanting to program too making A.I. like where to started and the steps afterwards. I see a bunch of videos and bootcamps but none of them tell what I should learn first they are all about what they are selling.

[@b7e286]

edge of my seat! very entertaining.

[@DoomRater]

"Nature just needed more time" Nature isn't going to get the blood vessels out of the way of my actual light receptors in my eyeballs!

[@uchihamadara6024]

Dude this series is amazing, would love to see some more stuff like this. also liked for ff7 music

[@spinningredchair8092]

In college i tried to design a little guy named Fredo, and his only task was to run around a town, pick up a block and carry it to the bloc bank in the middle, but during training for him, he got caught up in a loop that made him run in a circle 1 unit away from the block bank

[@MarexArtB]

Awesome series!

[@MetinFTW]

What program did you use to plot the graphs. Doesn't look like excel, but maybe I just don't know how to make it look as good as you do^^?

[@hjominous]

What do you use to create the courses?

[@eriko9]

what happened to your mousepad @ 8:14

[@esoartbook5955]

What software did u use for generating his course....?

[@shubhamdas9489]

What programming language do you use??

[@ArticulatelyFox]

Dude, first time watching your youtube. Hilarious brother. Thank you 😂

[@fivezeroniner]

what if you make/add a lot of forrests run in one go... like a competition... Will it make training faster? I know they also take computational power, But, will it be more efficient?

[@0przem]

lol the end segment is always so funny looking after watching you talk with your mouth closed all video

[@alexb9893]

Hi. I'm interested in what programming software you used as I'm planning on making a neural network for my A-Level Project and I can't think of anything better than C# or python (python I've never used before). Thanks

[@mskyba]

What is the name of the music that played while the Forrests ran, starting at 10:55? Amazing series btw, great work

[@Brokenrocktail]

Hey, question. What would happen if you put logical bounds on the output of the neural network. Like if you're about to be too close to a wall to avoid a collision, override the output and turn to avoid the collision, then give control back to the ANN when you're out of harms way? A hybrid approach. Use the neural network for navigation, but have an emergency hard-logic fallback for imminent collisions? Let me know what you think. Overall, great series. I subbed. Well worth watching, will recommend to friends.

[@ddayskeleton]

6:55 that is some risky stuff my friend, and I like that way of thinking

[@dylandowdy3687]

really cool stuff thanks for sharing with us

[@bigedwerd]

I believe you said the random/probability method worked out the best with training. Did you try this method multiple times and was it typically the best? If so have you considered seeding a random generator to keep the data consistent and tested it against other seeded random generators to see how they compared. Very cool series and channel.

[@iToastCrafter]

Really enjoyed this series :)

[@Aedi]

Nature doesn't use slice/probability, it uses random/probability, but the amount of randomness is obscenely tiny. Extremely small variations that occur very slowly.

[@AmCanTech]

What computer hardware are you using?

[@VojtechMach]

Can you explain why you chose such an architecture of the ANN? I mean I'd consider trying adding an extra hidden layer for complexity.

[@the_saucisson]

What, in your opinion, is the best way to start learning how to develop/create machine learning?

[@aar_ash]

most undreeated channel on youtube!

[@cough5431]

how'd you make those beautiful graphs? and what program do you use to train forest?

[@frenzy9741]

This man NEEDS SUBS!!!

[@theshortcut101]

I really liked that part at the end where we got to see you move your mouth! I know its weird but it just feels so much better to see your mouth move than not..... :P great videos btw

[@hanighate2096]

This is the first time i hear about these stuff but i think if u put a timer on forest ghosts(i call them that) and reward the one who runs the course in less amount of time it can be trained better! i luv the way u make these Videos keep up to good work :) Peace