video_id
stringlengths 11
11
| text
stringlengths 361
490
| start_second
int64 0
11.3k
| end_second
int64 18
11.3k
| url
stringlengths 48
52
| title
stringlengths 0
100
| thumbnail
stringlengths 0
52
|
|---|---|---|---|---|---|---|
O1b0cbgpRBw | like this is almost comically short so what you do is you put your state so s is your state okay s is it's the board as you have it right now okay this here that's this is s okay you put this into a neural network and the neural network gives you two things first of all it gives you p and and v so that's the second thing so v will simply give you a number v will tell you that this thing right here | 562 | 595 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=562s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | is about a plus 0.5 maybe so it says so plus one is winning and minus one is losing uh and it is this is called the value so maybe it says well this position i'm going to expect you to win uh roughly 75 of the time right which in expectation would be a value of positive 0.5 here because 75 percent of the time you win and the rest you lose let's say there is no draw in tic-tac-toe | 595 | 633 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=595s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | so there's this value function and the second thing is this p and the p is a policy function so the p will and i've drawn this a little bit maybe not super super duper too large but the p will tell you for every possible move you could make which one should you consider even okay so it maybe it assigns this here a point three and this here a point four but this here is like a point zero | 633 | 666 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=633s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | zero zero one and so on so for every possible move that you could do it will assign a number and it's a distribution so these numbers add up to one but that's not important it tells you which moves you should even consider going forward right so p in this case is a distribution over the next moves and with those two things together we can reduce our tree search quite a bit so now | 666 | 694 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=666s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | instead of expanding all the tree let's go back to the tree right here you can ask your p hey p which one of these three should i even consider and maybe p says you should only consider those two okay and then you go down and again you ask your p hey p which one should you consider and p maybe says well here you should consider those two here you should only consider that this | 694 | 721 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=694s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | one and this three over here we've we've already discarded this from the beginning okay so this p right here it guides your search it tells you at each point which moves should you consider and this as you can see reduces your tree dramatically in fact what alpha zero does is it simply says you have one second of time now expand as much as you can in this tree uh given this one second | 721 | 749 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=721s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | uh of of of time budget and the second thing is the value so what you would have to do expanding the tree is always to go to the end right so you always go to the end where at the end you have a fully filled board i don't know here x so you consider every possible situation okay here maybe this this player wins as you can see you always have to go to the end | 749 | 780 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=749s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | but in our case we don't want to always go to the end we'd rather explore more into like more branches than always go to the end and this is where the value comes in so at some point you simply say now i'm deep enough and now i'm going to ask my value v now there are slight differences with respect to alpha go and alpha 0 and so on but they all have in common that they | 780 | 808 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=780s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | estimate the value of the intermediate nodes using this v model from over here um i have v s v was green so they use this v model from over here to estimate at a certain depth so v learns to look into the future so everything that can happen from here and it estimates and it says well from here you maybe have a you know a 0.5 value or maybe a negative 0.7 and so on so v learns to assign these | 808 | 841 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=808s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | values to situations to states which are these nodes right here and p learns to suggest things to expand that's alpha zero and then at the end if you've expanded the tree enough and estimated well then you have a pretty good idea what's going to happen in each of the branches that you considered right in each of these branches you look into the future um from you here you look into the | 841 | 869 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=841s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | future here look into the future by doing this pv play and after one second after you've done you know a couple of hundred or thousand or however many uh looks into the future then you have a pretty good idea for each of the top level actions what's going to happen in the future and you can simply pick the one that has the best future for you according to your own model so that's | 869 | 897 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=869s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | what alpha zero does not so this is how you combine planning and neural networks you want to do planning but you can't because you can only go so deep so you use neural networks to first of all reduce the number of branches you consider because the neural network will tell you which ones are worthy to even look at and second of all you don't always have to plan to the end | 897 | 921 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=897s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | because you can simply ask your neural network how much an intermediate state is worth in expectation and this turns out to be pretty good why don't we do this for every single problem well we do for this we do need a simulator so you may recognize that right here i said we consider all the possible actions that we have and for each action we know exactly what's going to happen | 921 | 947 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=921s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | this is only possible like in a board game it's not even possible in like a board game where you have a a die to roll or a card to draw anything that is random there there is a way to include this right here but in this simple formulation we need to know exactly with 100 certainty what is going to happen if we take a particular action so this is only really applicable for | 947 | 972 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=947s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | the types of full information board games where we can write simulators that are pretty fast right and even then um even though chess you know has lots of available actions and complications it's nowhere near the complexity of like a let's say a modern video game or even or the real world is is completely out of scope for now for these types of things all right so that was alphago | 972 | 1,003 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=972s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | sorry alpha zero uh which builds on alphago of course and uh the rules of chess that we're going to consider using alpha zero are the following so there's no castling no castling for ten moves pawns can only move by one square forcing a stalemate is a win rather than a draw so you may know this in chess if you do not um checkmate the opponent's king but only | 1,003 | 1,033 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=1003s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | put them put the king in a situation where it cannot move that's called that's considered a draw and i think even in the chess community some people want to consider this a win there is torpedo where pawns can move by one or two squares anywhere on the board and semi torpedo where it's the same but only from the second and the third rank pawn back where pawns can move | 1,033 | 1,060 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=1033s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | backwards and pawn sideways where pawns can move laterally by one squares but captures are unchanged diagonally upwards and there is self capture where it's possible to capture one's own pieces so um there are you know slight slight details here with respect to the 50 move rule and so on but if you if you don't play chess simply consider these are changes minor in a lot of cases | 1,060 | 1,091 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=1060s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | minor changes to the chess rules that make the new rules either a superset or a subset of the original rules but they are going to have quite some changes in for the play and we're going to look at what happens so that's the entire research setup as you've seen it's alpha 0 applied to these new rule sets and under the assumption that alpha 0 will solve these will become master at these | 1,091 | 1,123 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=1091s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | games which we can't verify we can verify in chess because right alpha zero can beat people that have trained chess for all their life we can't verify it here so again this is an assumption so the first thing i want to look at here and this is going to play a little bit into my criticism of this paper it's a pretty cool paper but i do have some concerns right here | 1,123 | 1,149 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=1123s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | is the following uh the following charts so they do they do we don't consider how you train alpha zero let's just say you can train it you know to whatever pretty good performance here is how they evaluate so they evaluate for each variant they do 10 000 games played at one second per move for each different chess current so if you remember as we do our tree search right we expand | 1,149 | 1,183 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=1149s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | the tree according to our p and we estimate the values according to our v and we do this for one second in this first thing so in one second maybe this here is the tree so we have some sort of an understanding of what's going to happen in the future you can imagine if we have more time then we can expand this tree more and get a much more accurate picture of what | 1,183 | 1,209 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=1183s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | happens in the future okay so they do 10 000 games at one second per move but they also in addition to 1000 games played at one minute per move so there's 60 times more time and you can imagine that we'll add quite a number of nodes here and you know if if your p and v would be perfect then it wouldn't matter as much how much time you have as long as you sort of | 1,209 | 1,242 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=1209s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | have enough time but since they're not going to be perfect since they're only neural networks they're not uh god or schmidt hoover um they cannot accurately extremely accurately predict the future so this planning the the more you plan the more you actually look into the future the bigger your tree becomes the better moves you make so on the left you see the distributions | 1,242 | 1,268 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=1242s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | of winds losses and draws for one second per move and on the right for one minute per move so both white and black pieces here are played by alpha zero so it's not alpha zero against something else this is playing against itself and you can see in uh in classic chess it's it's quite it's quite saddening actually um that this game which is is so famous you can see that | 1,268 | 1,299 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=1268s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | in of 10 000 plays 8 820 end in a draw which means that if both players are super duper good and uh and and play you know play against each other it most likely is going to be a draw and this i think is the criticism even in human chess is that it's not really a decisive game in that it ends a lot of times in a draw so one of the motivations here would be can | 1,299 | 1,332 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=1299s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | we find a rule set that is maybe more decisive so that's one of the investigations they do in the paper but you can see that there are actually so if you consider this torpedo chess right here um there it is more decisive as you can see in more times either white or black winds right here um and there are others which are even less decisive like pawn back so | 1,332 | 1,360 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=1332s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | when pawns can move back then uh players may just camp they like move a pawn forward and move it back again and that will lead to a lot of closed plays and so on whereas torpedo makes you move much faster you can advance your pawns much faster and that will probably lead to the end much faster so if you consider this on the right so what changed the rules didn't change | 1,360 | 1,387 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=1360s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | alpha 0 didn't change it simply changed that we now let alpha 0 think for longer and you can see that the decisiveness reduces dramatically so whereas 88 resulted in a draw with one second per move now 98 result in a draw with one minute per move and this is a trend throughout these games and that's also what they say in the text it is to assume that if you let alpha zero plan for even | 1,387 | 1,422 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=1387s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | longer that this trend will continue and ultimately whatever rule set you make the result is going to be a draw um if two two let's say perfect players play against each other which is a bit which is a bit saddening right because um yeah that ultimately ultimately means that all of these rules aren't decisive it's only they're only decisive due to the fact that either | 1,422 | 1,454 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=1422s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | um one or the other players is way better or or that in general that they are not they are not perfect um which is an appeal of the game but there are certainly games that are decisive even though both players are pretty high level i mean think of every every competitive video game um so yes so that's a bit of my criticism all of this all of this needs to be analyzed in the background | 1,454 | 1,485 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=1454s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | that what's actually happening here is that we're dealing with imperfect decision making due to a limit in resources okay and this assumption now is already a little bit invalid right the assumption we made at the beginning why i pointed this out is that alpha zero can solve these games let's say to perfection and here when we analyze the decisiveness and so on it seems to be | 1,485 | 1,511 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=1485s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | purely or largely a factor of um how much time alpha zero has to think about the moves and these two things to me they don't really go go together because we don't know if for a different rule set um you know the training is harder or might take longer and so on or that this exact one second makes a difference or not it's it's just um there are so many variables here and | 1,511 | 1,544 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=1511s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | when you're dealing with let's say imperfect systems that are not trained to the end or evaluated in their full potential you're always dealing with the fact that you stopped each thing at some intermediate point and that intermediate where that intermediate point is can influence the results drastically now here it seems at least the ordering isn't changed by much but um | 1,544 | 1,572 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=1544s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | yeah this is one let's say one criticism the other criticism here uh that that i would have again is the fact that if you consider something like torpedo where you can move much much faster then yes of course uh let's say i don't know is it more interesting that's that's the question right here so they look at a lot of things like decisiveness diversity and so on | 1,572 | 1,600 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=1572s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | but the question is is it more or less interesting to play and i think that's what humans are really after and they're sort of trying to find proxies to this um i would argue if you play something like torpedo the game's maybe much faster and um so you you get to the end faster but also maybe might not be as interesting even though it's it's faster uh because | 1,600 | 1,625 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=1600s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | your the complexity is is less and with respect to the decisiveness here so if you have a game that's faster um you also need to take this to into account because here is another thing that is sort of an arbitrary choice as moves are determined in a deterministic fashion given the same condition diversity was enforced by sampling the first 20 plies in each game proportional | 1,625 | 1,656 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=1625s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | to their mcts visit counts so what does that mean that means that if you run alpha 0 on the same situation on the same tree sorry on the same board position it will always come up with the same move except for parallelism inconsistencies and so on but it will in you know in in a lot of times it will come up with the same move so how do you play 10 000 games because you can just play one | 1,656 | 1,687 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=1656s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | game because each game will be the same because you simply tell alpha zero give me your best move right so it will just play its optimal strategy and all the games will be exactly the same so there's no reason why these should come out different so they enforce diversity by saying okay okay in the first 20 moves of a game we don't actually take the best move | 1,687 | 1,710 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=1687s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | right usually you have you have this distribution at the end of the tree search you have a distribution where you say okay this move right here is clearly the best move i'm going to play this however if this is one of the first 20 moves of the game they say no we need a bit of diversity uh so we're going to sample according to this distribution rather than just play the best one | 1,710 | 1,735 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=1710s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | now this number 20. it's just sort of decided arbitrary right and if you consider something like torpedo it's a faster game so you're faster in opening faster make and you're faster to the end game maybe even though they say well the game length isn't affected this much it could just be that um you're faster in a situation where um you're kind of forced to do | 1,735 | 1,764 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=1735s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | certain moves and maybe the difference in decisiveness here is simply a result of the combination of the faster uh moves in torpedo together with this the fact that they just keep the 20 plies for each game again this is something that you need to consider when analyzing these results right here and there are a number of these choices um right here like the one second or one | 1,764 | 1,794 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=1764s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | minute per move we sample for the first 20 plies before we play the max move that where i think the the results of the study right here they have rather limited interpretability if you if you ask me because um because of these of these choices now of course they're still the results are quite plausible believable and the idea is really cool to explore these rule sets but this | 1,794 | 1,824 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=1794s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | was this is just my criticism right here so we'll go through the rest of the results pretty pretty quickly because a lot of people aren't chess enthusiasts and we'll just pick out kind of the core messages that the paper is trying to get across so here the table again with respect to decisiveness and you can see even uh for so for classic chess it's a white has a 50 this is the empirical | 1,824 | 1,854 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=1824s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | score for white under different game conditions so 50.8 percent means most of the time it's a draw so white wins uh with a probability of 50.8 uh most of the time it's a draw and you see even like the most decisive variant torpedo right here is a 54 only um so they they analyze different defenses and how the decisiveness is with respect to different defenses that are not really popular | 1,854 | 1,888 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=1854s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | under classical chess and the results are interesting if you play chess but i would say they're rather they're kind of aha okay if you do not play chess because they consider individual moves and so on what is an interesting part is um this right here where they look at they look at one move that in classical chess so e4 is a very very um popular opening where you move your e | 1,888 | 1,922 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=1888s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | pawn twice for white and nf3 is not a super popular opening and here they compare this in classic chess and in no castling gesso this thing right here is a histogram and the histogram shows you the log probability of opening sequences when you play the individual moves so what does this mean right here if you play e4 then the distribution is something like this which means that | 1,922 | 1,961 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=1922s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | you have some sequences that have no entropy at all which means that once you play e4 and maybe one move more then it's almost it's almost determined what you have to do according to alpha zero you have like no choice except play these few next moves um however if you play nf3 then alpha zero says look this distribution is much more to the right which means that you have a lot more options here | 1,961 | 1,995 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=1961s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | now again this could be because the move is actually less decisive because the move leads to more balanced more interesting situations where you can continue however you know with many choices it could also be because it's simply alpha zero simply doesn't know as well what to do because it leads to more complicated games and you get to give each move one minute to evaluate alpha zero might | 1,995 | 2,021 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=1995s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | just not be as good in those situations because it leads to more complicated situations if it could search for longer maybe this distribution would shift over here just as well again we don't know because you only give this one second or one minute each time for both um and again this goes under the assumption of alpha zero as this perfect player however back to what they | 2,021 | 2,048 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=2021s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | want to say here if you do this in no castling chess you can see that uh this spike right here are all the these berlin defense variants and castling this 0 right here is a big part of that line if you do this in no castling chest you can see that these two moves now the histograms overlap much more which means that and in fact you can see in the in this number of possible moves right | 2,048 | 2,076 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=2048s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | here that they come closer together so not only does the blue shift to the right the orange actually shifts to the left and it basically means that whether you open with e4 or knight f f3 you are going to have about the same complexity of game the same number of moves available to you going from there as you can see right here these lines are the moves available for white and black | 2,076 | 2,105 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=2076s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | under the different rule sets so in e4 here especially as black you do not have many moves available as white a little bit more but also not more um whereas in no castling you do so again small rule change uh big effect on the possible moves that you have can consider and this is the type of this is the type of information that you would want to have when you | 2,105 | 2,137 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=2105s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | design a game and they allude to this also at the end here in their conclusions so the last thing is they also compare the material values of the pieces here in the different rule sets as you might imagine so some pieces become much more or less valuable i find it particularly interesting that if you do something like pawn sideways or then where the pawns are much more | 2,137 | 2,166 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=2137s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | powerful of course all the other pieces drop in value again these results are pretty plausible so i don't want to trash the paper right here because it seems like it seems like the the results are as i say plausible and can give some cool insights so the chess master also gives um gives his opinions on these different strategies that alpha zero comes up with | 2,166 | 2,193 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=2166s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | for the different rules and let's go through the conclusions real quickly so they say assessing the consequences of rule change in the game design process demonstrate on chess where we've trained alpha zero to evaluate nine different variants representing atomic changes to the rules of the game training alpha zero model on these rules changes helps us effectively simulate decades of human | 2,193 | 2,218 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=2193s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | play in a matter of hours and answer the what if question what the play would potentially look like under developed theory in each chess variant we believe that a similar approach could be used for auto balancing game mechanics in other types of games including computer games in cases when a sufficiently performant reinforcement learning system is available and yes this is | 2,218 | 2,242 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=2218s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | i mean this the application here would be for something like this if you design a new game then you want to know what you have some choice with how you can make the rules and you don't want to let humans become really good at each of the rules and then compare you can simply give this to the algorithm and the algorithm will tell you what kind of plays result from each | 2,242 | 2,266 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=2242s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | rule set and then you can choose the one that you find most interesting or most uh maybe commercially viable and what not i actually see this much i see this bigger than just games and this alludes a bit to the salesforce paper on this ai economist i think we can let ai you know get tell us what happens if we change for example uh things like tax policy or any any sort of policy i know | 2,266 | 2,299 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=2266s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | humanity is very complex to model and so on and you're never going to have a perfect simulator which probably makes alpha zero not good but in limited situations like maybe also stock trading rules and so on you could definitely have situations where the rule set is too complicated to solve analytically but you could give it to an rl algorithm and see what happens and whether or not you like | 2,299 | 2,326 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=2299s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | the outcome and whether or not there are any like obvious exploits that uh you did not see so this i find you know pretty it's it's a pretty cool approach and and we should think of this in the future as we build systems that have rules in whatever capacity be this games or policy so the they say okay yada yada yada we showed that there are several chess variants among those | 2,326 | 2,355 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=2326s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | considering the study that are even more decisive than classical chess meaning torpedo chess semia torpedo chess no castling chess and stalemate equals winches we quantified a rising diversity of opening play and the intersection of opening trees between chess variations showing how different the opening theory is for each of the rule changes yeah they again this this diversity | 2,355 | 2,380 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=2355s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | of opening play it really rests on this assumption that alpha zero is a is a good player and any sort of an equally good player in all of these variants right because if it's worse in a variant it might not be as sure about the moves and that would just look like oh you have many possibilities but in fact alpha zero is just worse at it and it doesn't know so they also look at | 2,380 | 2,405 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=2380s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | the intersection of opening trees like if you change a rule how does this change um change the the kind of how does this change the the initial game so a lot of these grandmasters they learn by heart all of these opening trees the initial moves of a game how much would they have to relearn there is a negative correlation between the overall opening diversity | 2,405 | 2,429 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=2405s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | and decisiveness as decisive variants likely require more precise play with fewer plausible choices per move again this is one view right the other view is that there are rule sets that are just make it into a harder game and then alpha zero given the same amount of compute is a worse player and therefore it can't play as well therefore the games are less decisive | 2,429 | 2,461 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=2429s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | and also the opening diversity is higher because it doesn't know if the game could be as decisive it might just be an effect of alpha zero for each of the chess variants we estimated yada yada okay no castling chess being the first variant that we analyzed has already been tried in experimental blitz grand master tournament in chennai as well as a couple of longer grand | 2,461 | 2,486 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=2461s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | master games our assessment suggests that several of the assessed chess variants might be quite appealing to interested players and we hope that this study will prove to be a valuable resource for the wider chess community i yeah i don't know is is the chess community flourishing or going under recently because it seems to me like it once once a game is solved that hard | 2,486 | 2,511 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=2486s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
O1b0cbgpRBw | by computers i mean it's still fun but um yeah i just i just i guess counter strike is also solved by bots real hard uh it's just impressive when humans play or so um yeah i don't know all of this is again if you're into chess look into this paper they have a lot of really interesting results that are not interesting to go into for the general community but i | 2,511 | 2,541 | https://www.youtube.com/watch?v=O1b0cbgpRBw&t=2511s | Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess (Paper Explained) | |
lbKg3OSTsgA | [Music] this video will present a recent paper from Google AI revisiting self supervised visual representation learning so the headline idea this paper is that the standard architecture designs and convolutional neural network advances that have been working in supervised learning tasks like image net classification don't necessarily translate to these self supervised tasks | 0 | 24 | https://www.youtube.com/watch?v=lbKg3OSTsgA&t=0s | Self-Supervised Learning | |
lbKg3OSTsgA | such as predicting the rotation the permutation of jigsaw puzzles or the exemplar augmentation task so this is the ideas are the neural network architecture designs over fitted to sell to supervised learning tasks such as recognition and detection and should maybe these neural architecture searches be deployed into self supervised learning or into like a pipeline of self | 24 | 46 | https://www.youtube.com/watch?v=lbKg3OSTsgA&t=24s | Self-Supervised Learning | |
lbKg3OSTsgA | supervised learning and then taking the representations into the classification models but then using the hyper parameter neural architecture search and all these heuristic tricks to design this jointly on self supervised tasks forth for the downstream representation learning so self supervised learning is inspired by NLP success and it's just recently it's been tested since like the | 46 | 70 | https://www.youtube.com/watch?v=lbKg3OSTsgA&t=46s | Self-Supervised Learning | |
lbKg3OSTsgA | dawn of alex net but it is definitely gaining traction and so an NLP this is about predicting words from their context so when the sentence predict words from their context words and form would be labelled as positive words and predict one and then tiger an ocean don't appear in the context so they've been labeled as zero or negative so the sole supervised learning has these | 70 | 90 | https://www.youtube.com/watch?v=lbKg3OSTsgA&t=70s | Self-Supervised Learning | |
lbKg3OSTsgA | pretext tasks and so there's other techniques other than context like NLP there's more ideas in visual representation learning and images and computer vision so here the some of the common self supervised visual learning tasks rotation prediction example are classes relative patch location and jigsaw puzzle permutations and many of the studies that have already come out | 90 | 114 | https://www.youtube.com/watch?v=lbKg3OSTsgA&t=90s | Self-Supervised Learning | |
lbKg3OSTsgA | on this use the Aleks net CNN architecture this paper from Google AI is going to use the state-of-the-art ResNet designs like the wide ResNet and then a reversible resume which is a more efficient implementation so rotation task is like this you would take images and then you rotate them either zero degrees 90 degrees 100 to 70 and then the network is basically | 114 | 136 | https://www.youtube.com/watch?v=lbKg3OSTsgA&t=114s | Self-Supervised Learning | |
lbKg3OSTsgA | doing a for class classification task for these different rotations the exemplars used data augmentations to take an image and then change it in like a ton of different ways and then this image corresponds to its own class so it would be like a massive way like maybe like a thousand class and on classification problem so jigsaw puzzle this is an interesting one it's where | 136 | 160 | https://www.youtube.com/watch?v=lbKg3OSTsgA&t=136s | Self-Supervised Learning | |
lbKg3OSTsgA | you take like a patch of the image and then you like you crop the image and then you further refine the crops and then scrambled away in this thing and then try to have you pass each of the crops through the network so each of these each of these square tiles goes into a convolutional Network and then it predicts the where it thinks it might lie in the permutation so I personally | 160 | 183 | https://www.youtube.com/watch?v=lbKg3OSTsgA&t=160s | Self-Supervised Learning | |
lbKg3OSTsgA | don't I'm not really a huge fan of this task I don't really see how it makes a whole lot of sense because how does it know the context really and then relative patch location is the same idea I like this idea more where you just have two patches and you predict how they might relate to one another and then another really interesting study that I'm gonna be making a video on | 183 | 203 | https://www.youtube.com/watch?v=lbKg3OSTsgA&t=183s | Self-Supervised Learning | |
lbKg3OSTsgA | tomorrow so please subscribe if you're interested in this video is a multitask table supervised visual learning where you combine these self supervised learning tasks together and then see what kind of representations are derived from that and they test this and their paper using the ResNet 101 model so representations for image classification what you do is you freeze most of the | 203 | 225 | https://www.youtube.com/watch?v=lbKg3OSTsgA&t=203s | Self-Supervised Learning | |
lbKg3OSTsgA | network and then you take the pre largest layer like the like an intermediate vector representation and then you input that to a logistic regression model and use that as a classifier train with SGD and data augmentation so the vector that they extract feature vector from the sub supervised learning task they vary this from size 2048 4096 6000 144 and 8192 so | 225 | 251 | https://www.youtube.com/watch?v=lbKg3OSTsgA&t=225s | Self-Supervised Learning | |
lbKg3OSTsgA | this is the size of the representation of vector extract from the self supervised learning task so another thing they find is you might ask like how about instead of a logistic regression we do a more complex multi-layer perceptron model so this plot just shows this is the logistic regression in the MLP and they basically perform the same so the logistic regression is has sufficient | 251 | 272 | https://www.youtube.com/watch?v=lbKg3OSTsgA&t=251s | Self-Supervised Learning | |
lbKg3OSTsgA | capacity for this and then another interesting thing is we're in the network do you get the features from so this one thing that's interesting is in the VG g19 network they take the intermediate features from the third block rather than the very end and get better results but with the ResNet architectures they always get better results towards taking it as close to | 272 | 294 | https://www.youtube.com/watch?v=lbKg3OSTsgA&t=272s | Self-Supervised Learning | |
lbKg3OSTsgA | the output as you can get so again this would mean like should you take it from like here or should you take it from down here like towards the output or from the intermediate features so the data sets to the tests our image net which is 1.3 million images in the thousand classes and then the places 205 which is 2.5 million images in 205 classes and these are pretty different | 294 | 318 | https://www.youtube.com/watch?v=lbKg3OSTsgA&t=294s | Self-Supervised Learning | |
lbKg3OSTsgA | data sets and it's used just for a measure of generalization with respect to the data sets so these are the results basically across the rev net ResNet v2 ResNet v1 just showing how amazingly different the results can be for the different architectures so and not only that but inconsistent so even though Greg net kills on rotation it doesn't perform as well on a relative | 318 | 344 | https://www.youtube.com/watch?v=lbKg3OSTsgA&t=318s | Self-Supervised Learning | |
lbKg3OSTsgA | patch location so this is the results in the table so this refers to increasing the widening factor of a resinate so increasing the number of feature maps to each intermediate layer so very interesting trend you see is as they continue to increase the capacity of the model they get better results so the widening factor is highly correlated with success on this task and yeah so as | 344 | 370 | https://www.youtube.com/watch?v=lbKg3OSTsgA&t=344s | Self-Supervised Learning | |
lbKg3OSTsgA | they increase the representation capacity of the model Bay get better results so then also this shows how applying their new ResNet with the winding factor how this compares to the previous papers that have been published on self supervised learning so most interestingly the rotation paper that first came out using an Alex net style architecture achieves 38.7% image net | 370 | 391 | https://www.youtube.com/watch?v=lbKg3OSTsgA&t=370s | Self-Supervised Learning | |
lbKg3OSTsgA | but their model achieves like 20% better and is getting much closer to the fully supervised benchmarks so this is another interesting thing that they present is that the success on this Hope supervised task isn't always correlated with image accuracy like this point right here has like 95 percent on rotation but then only like 20 percent on that image net accuracy so then | 391 | 417 | https://www.youtube.com/watch?v=lbKg3OSTsgA&t=391s | Self-Supervised Learning | |
lbKg3OSTsgA | another very interesting takeaway from the study is that it seems like larger models like increasing the width multiplier and this is the vector that you take from the representation seems like as you increase the width the network and then as you get a bigger feature from the model that you do a better job on image met classification with these vectors so thanks for | 417 | 440 | https://www.youtube.com/watch?v=lbKg3OSTsgA&t=417s | Self-Supervised Learning | |
xtOg44r6dsE | [Music] hi guys this is Laker from a Judaica the evolution of AI has changed the entire 21st century in terms of technology ai has told in the spotlight and it's advancements are quicker than we predicted with such an exponential growth in AI machine learning is becoming the most training field of the 21st century it is starting to redefine the way we live and it's time we | 0 | 34 | https://www.youtube.com/watch?v=xtOg44r6dsE&t=0s | Supervised vs Unsupervised vs Reinforcement Learning | Data Science Certification Training | Edureka | |
xtOg44r6dsE | understood what it is and why it matters in this session we'll be discussing the different types of machine learning and we'll compare them to each other so let me run you through today's agenda we're going to begin the session with an introduction to machine learning next we will discuss the types of machine learning after that we'll compare supervised unsupervised and reinforcement learning | 34 | 55 | https://www.youtube.com/watch?v=xtOg44r6dsE&t=34s | Supervised vs Unsupervised vs Reinforcement Learning | Data Science Certification Training | Edureka | |
xtOg44r6dsE | based on a few key parameters we'll finally end the session by discussing a few example problems that can be solved using supervised unsupervised and reinforcement learning algorithms so without any further delay let's get started so guys machine learning is the science of getting computers to act by feeding them data and letting them learn a few tricks on their own without being | 55 | 79 | https://www.youtube.com/watch?v=xtOg44r6dsE&t=55s | Supervised vs Unsupervised vs Reinforcement Learning | Data Science Certification Training | Edureka | |
xtOg44r6dsE | explicitly programmed now this sounds awfully a lot like a human child so let's consider a small scenario to understand machine learning now as a child if you had to distinguish between fruits such as cherries apples and oranges you wouldn't even know where to start because you're not familiar with how the fruits look now as we grow up we collect more information and start | 79 | 102 | https://www.youtube.com/watch?v=xtOg44r6dsE&t=79s | Supervised vs Unsupervised vs Reinforcement Learning | Data Science Certification Training | Edureka | |
xtOg44r6dsE | developing the capability to distinguish between various fruits the only reason why we are able to make this distinction is because we absorb our surroundings we gathered more data and we learn from our past experiences it's because our brain is capable enough to think and make decisions since we have been feeding it a lot of data and this is exactly how machine learning works it involves | 102 | 125 | https://www.youtube.com/watch?v=xtOg44r6dsE&t=102s | Supervised vs Unsupervised vs Reinforcement Learning | Data Science Certification Training | Edureka | |
xtOg44r6dsE | continuously feeding data to a machine so that it can interpret this data understand the useful insides detect patterns and ident my key features to solve problems this is very similar to how our brain works now let's move ahead and take a look at the different types of machine learning so first of all we have supervised learning now guys supervised means to oversee or | 125 | 148 | https://www.youtube.com/watch?v=xtOg44r6dsE&t=125s | Supervised vs Unsupervised vs Reinforcement Learning | Data Science Certification Training | Edureka | |
xtOg44r6dsE | direct a certain activity and make sure it's done correctly in this type of learning the machine learns under guidance so at school or teachers guided us and taught us similarly in supervised learning machines learn by feeding them label data and explicitly telling them hey this is the input and this is exactly how the output must look okay so the teacher in this case is the training | 148 | 172 | https://www.youtube.com/watch?v=xtOg44r6dsE&t=148s | Supervised vs Unsupervised vs Reinforcement Learning | Data Science Certification Training | Edureka | |
xtOg44r6dsE | data next we have unsupervised learning unsupervised means to act without anyone's supervision or without anybody's direction now here the data is not labeled there is no guide and the machine has to figure out the data set given and it has to find hidden patterns in order to make predictions about the output an example of unsupervised learning is an adult like you and me we | 172 | 197 | https://www.youtube.com/watch?v=xtOg44r6dsE&t=172s | Supervised vs Unsupervised vs Reinforcement Learning | Data Science Certification Training | Edureka | |
xtOg44r6dsE | don't need a guide to help us with our daily activities we can figure things out on our own without any supervision finally we have reinforcement learning now guys reinforcement means to establish or encourage a pattern of behavior let's say that you were dropped off at an isolated island what would you do now initially you'd panic and you'd be unsure of what to do where to get | 197 | 220 | https://www.youtube.com/watch?v=xtOg44r6dsE&t=197s | Supervised vs Unsupervised vs Reinforcement Learning | Data Science Certification Training | Edureka | |
xtOg44r6dsE | food from how to live and so on but after a while you will have to adapt you must learn how to live in the island adapt to the changing climates learn more to eat and what not to eat so here you're basically following the hit and trial concept because you new to the surrounding and the only way to learn is experience and then learn from your experience this is what reinforcement | 220 | 244 | https://www.youtube.com/watch?v=xtOg44r6dsE&t=220s | Supervised vs Unsupervised vs Reinforcement Learning | Data Science Certification Training | Edureka | |
xtOg44r6dsE | learning is it is a learning method wherein an agent which is basically you stuck on the island interacts with its environment which is the island by producing actions and discovers errors or rewards and once the agent gets trained it gets ready to predict the new data presented to it now let's move ahead and look at the differences between supervised answer | 244 | 267 | https://www.youtube.com/watch?v=xtOg44r6dsE&t=244s | Supervised vs Unsupervised vs Reinforcement Learning | Data Science Certification Training | Edureka | |
xtOg44r6dsE | and reinforcement learning so let's begin by looking at their definitions now like I mentioned earlier supervised learning is a type of machine learning wherein we teach the machine using label data so an input and your output is label next we have unsupervised learning over here the data provided to the machine is not labeled and the machine has to learn without any supervision so | 267 | 291 | https://www.youtube.com/watch?v=xtOg44r6dsE&t=267s | Supervised vs Unsupervised vs Reinforcement Learning | Data Science Certification Training | Edureka | |
xtOg44r6dsE | that's why it should discover hidden patterns and trends in the data finally we have reinforcement learning now the basic concept behind reinforcement learning is that there is an agent now this agent is put in an unknown environment so the agent has to explore the environment by taking actions and transitioning from one state to the other so that he can get maximum | 291 | 313 | https://www.youtube.com/watch?v=xtOg44r6dsE&t=291s | Supervised vs Unsupervised vs Reinforcement Learning | Data Science Certification Training | Edureka |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.