Leveling up Nana's 'Reference AI'

[TomoYoMomo]

Wow that pretty crazy that something so simple is only being discovered now and it has a huge impact on the IC meta entirely

 

[Nintendude]

I'll try doing this before playing matches from now on and see if I notice anything. The clips cited as "evidence" really don't indicate anything to me since Nana always does random good things once in awhile.

One question I have though is that wouldn't a level 9 Ice Climbers have a better Nana than a level 1 Ice Climbers if this was in fact true?

 

[jdmcwombat]

the other CPU modes don't have any code that make the characters use any Side-B moves to recover.

Luigi?

 

[LydianAlchemist]

Luigi?

lets be honest when has that ever worked heh

 

[Fly_Amanita]

I'm not remotely memory address-savvy, but this seems potentially noteworthy to me.

 

[Scidadle]

Omg this is amazing.

Good work finding this

 

[Fly_Amanita]

Some updates:

I repeated the test I mentioned near the end of the last page a bunch of times and kept track of whether Nana would hit the ceiling and land near the bottom area again or would fly into the tunnel in the upper right of the bottom area.

With a "level 1 AI", I did this 20 times. She bounced into the ceiling 16 times and went into the corridor 4 times. Perhaps significantly, the score between the options was initially 3-3, and then there was a long string of ceiling hits, punctuated by one other corridor entrance after the 13th ceiling bounce.

With a "level 9 AI", I initially did this 20 times. She bounced into the ceiling 9 times and went into the corridor 11 times. However, I suspected that a slight detail of how I was setting things up might be significant (whether I crouched before or after the fsmash charge began), and did some more after that (all in the same game). My suspicion didn't seem to be the case, but what I observed was still weird. I did have the ICs crouch after the fsmash charge began (the opposite of what I was mostly doing prior to this) and 6 out of 6 times, Nana did bounce off the ceiling. However, after that, out of the 6 times I had Nana crouch before the fsmash charge began (what I was mostly doing the first in the first 20 iterations of this part), she still bounced off the ceiling 4 times. If we add these to the total thus far, that would bring it to 19 ceiling bounces and 13 corridor entrances.

It does seem quite possible to me that there is some other variable floating around (possibly one effectively irrelevant to what is happening in-game) affecting the odds. It did feel to me like there was a trend for one option to occur over and over with only brief appearances by the other; I'm aware stuff like this tends to appear in long strings of chance events, but it seemed extreme enough to me to be conceivably attributable to something else looming about. I can't confidently assert anything thus far.

 

[guedes the brawler]

for reference, does this work in brawl too?

because if it does we might be seeing this for Nana and Luma in smash 4, most likely

 

[Salem Beats]

Pretty crazy.

Now the question:
Will it be banned?

 

[Chemlo]

It's cool how this thread starts with a video of me doing something dumb against Ice Climbers.

 

[kelots]

any difference with a level 5 in that experiment fly?

 

[R0Y]

13 years later, damn.

 

[Fly_Amanita]

I've only looked at levels 1 and 9 so far. I can look at other levels later, although I'd rather at least get something more conclusive with the two I'm looking at so far before then.

 

[kelots]

reason i ask is that my understanding from boards on practicing combo's/tech is that level 1's dont DI, level 9's di inwards but the 3-7 range have random (or maybe set but mroe varied?) DI - which is what i think is happening in Myougi's video

my understanding is often wrong, i'll see if i can spot anything tonight and post results

 

[snaprocks]

20XX and we still finding glitches I love this.

 

[Bones0]

I don't see any way this will end up being legal in tourney, but it's certainly fascinating if it's true. It reminded me of that event match where you have to only KO the Nanas. Maybe that could possibly be used for more ideal testing somehow (just throwing it out there).

 

[ResidentWaffle]

I'm not remotely memory address-savvy, but this seems potentially noteworthy to me.

According to this it seems unlikely.

In the past I did some AI manipulation in training mode, and have messed around with memory addresses in dolphin and from my experience usually when I modify a memory address(as is the case usually in assembly) other functions override it and make to go back to its normal value. Often values like these are dead values that just are placeholders to get overridden. Furthermore, if it is the case that the AI level changes throughout the match it would make sense that it would be due to random seeds choosing a level and to me it seems highly unlikely that these seeds would be influenced by AI levels selected beforehand(although still possible, but also probably almost statistically insignificant if it is).

Edit: Can you give us the memory addresses you used to find your correlation, myougi?

 

[Myougi]

Can you give us the memory addresses you used to find your correlation, myougi?

Aliami was the discoverer of the trick and knows all the memory addresses but I think this may be what you are looking for. Aliami put some stuff in his post where we shows them changing in a video.

The memory address for Nana's CPU mode is 0x0048082f.

It does seem quite possible to me that there is some other variable floating around (possibly one effectively irrelevant to what is happening in-game) affecting the odds.

Could percentage be the other variable?

There may be some kind of 'trigger' that activates her aggressive side when a certain percent is met. I'll look into this myself, but just reading the thread and going back over some personal instances leads me to believe that this is not the case.

 

[FlamingForce]

Could percentage be the other variable?

I distinctly recall Leffen telling me that Nana does not use smash attacks as long as Popo is below 50%, I believe I even made a thread about it. Other then that I have no clue.

 

[Fly_Amanita]

Could percentage be the other variable?

Popo was always at the same percentage when I did this, namely, 100%. I had both characters SD after each iteration of the test so that ICs always started at 0% and Fox's lasers were unstaled at the start. I'd think Nana would also have been hit by all the lasers, too, but it's possible that that's not the case. I did notice when doing stuff with Samus that her charge shot would sometimes only hit Popo and sometimes hit both ICs, seemingly depending on what part of the idle animation ICs were in, but that was with ICs standing and not crouching. ICs sway around considerably less while crouching, which is why I had them do that when getting hit by Fox's lasers. Overall, I'm doubtful that percentage variations are a factor in what I observed.

 

[TheReflexWonder]

Seriously, someone should just trying shooting a wayward Nana with projectiles. She'll powershield one pretty quickly if she's level 9, won't she?

 

[Fly_Amanita]

One of the first things I did was have a Falco shoot lasers across Hyrule at a stray Nana. Nana never powershielded any of them, regardless of level or percentage (I looked at 0% and something like 60% and levels 1 and 9, IIRC). She might need to have a target in mind already to do such things, though.

 

[FlamingForce]

The creator of the reddit thread noted in one of his posts that the programming of the "NANA" AI references the CPU level only when she

1. DI's
2. Defends herself.

I'm not entirely sure how you tested her powershielding but if she were trying to get back to Popo during the laser hits, its that she simply wasn't in "defend yourself" mode.
Nana does have a way of switching between running back to popo the one time or deciding to stand her ground and attack the other time (And usually failing spectacularly)
The latter might see improvement through this trick though I don't know when she goes for one or the other.

 

[Engo]

While I agree it's important to prove whether this as a pseudo effect or a genuine difference in AI, if they're so hard and ambiguous to differentiate then this probably isn't that great of an advantage for IC players anyways...

 

[Nintendude]

I played about 30 minutes of friendlies with CTL using a "level 9" Nana. While she did a few wonky things during that session (an f-smash edgeguard and also falling down with a fair), it was nothing that was implausible for Nana to do under normal circumstances. I also switched back to a "level 1" Nana and didn't notice any differences in her behavior after switching.

My gut feeling is that this trick doesn't actually work and that the "evidence" is a combination of confirmation bias and people just paying more attention to Nana's behavior, leading them to noticing things they normally wouldn't notice. I'll report back if I notice anything that leads me think otherwise.

 

[Unlimited]

In my test sthere is no way this works, Nanas has acted the same in all of my testing.

It's easy to desync the ICs. Grab ledge with Popo, get you friend to move her to the other side of the stage. She wont run toward you anymore at this point. I tried it at least 10 times. She stood there as if she was a level 1 computer almost every time. After I did this I tried it without doing the level 9 trick. She did the exact same thing. Nana has ledgeguarded and done lots of fancy stuff for me in the past. I'd say that those videos just were by chance.

 

[Fly_Amanita]

I'm also inclined to think confirmation bias is a big part of it. In general, I'd think that if people expect something to be different from how it used to be and there isn't much dire need to know if there really is a difference, people will be inclined to see one. I'm reminded of that time when Amazing Ampharos was showing off a build of Balanced Brawl to some good players; the players all felt that Olimar would be the new undisputed best because of how strong his new usmash was, but the move was, in fact, unchanged.

 

[TheReflexWonder]

Reminds me of how there was a long debate about whether or not Brawl CPUs learned from human play.

 

[Fly_Amanita]

Just for the sake of having data on video, I'm going to be repeating the test I did earlier to gather more substantial data. It'll be incredibly boring to watch if anybody decides to, but I just want it documented: http://hitbox.tv/flyamanita

 

[TobiasXK]

Edit: Video's done. Took me a minute or two to do with SnagIt.

this is not useful in terms of proving anything about in-game behavior; you would have to show observable differences in whether/how the NANA AI references the stored CPUlvl value while desynched.

 

[Fly_Amanita]

So, I streamed the following:

I did twenty iterations of the Hyrule test in four games each. The first and fourth games were with a "level 9 AI" and the second and third were with a "level 1 AI". In the first game (with a level 9 AI), Nana bounced into the ceiling 14 times and entered the tunnel in the upper right 6 times. In the second game (with a level 1 AI), Nana bounced into the ceiling 15 times and entered the tunnel 5 times. In the third game (with a level 1 AI), Nana bounced into the ceiling 10 times and entered the tunnel 10 times. In the fourth game (with a level 9 AI), Nana bounced into the ceiling 11 times and entered the tunnel 9 times. If people care about the order in which the data occurred, I have that available in this pastebin, (games listed from top to bottom, individual iterations of the test from left to right). The variation within the tests of the same level AIs makes me very dubious that any differences observed are due to AI differences in the first place.

Short Fuse noted a correlation near the end that Nana's DI appeared to depend on the percentage the fsmash put Popo at (either 120% or 121%), which I can look at when reviewing the archives. If the DI is consistent with Popo's end percentage, then the AI still difference still wouldn't appear to matter.

 

[S2rulL]

Could there be a possibility that Nana's AI cycles each time you play with the Icies, as opposed to it being selectable? Or could there be some element of RNG to it, that it just assigns you an AI level for Nana and the chances of getting a lower level is just higher?

 

[TryHardOliver]

COOL THANKS!!

 

[Encephalon Decay]

So I'm just cherrypicking at this point for the sake of keeping interest in Nana's AI, but is it perhaps possible that Nana's AI level is influenced by Popo's %? Said theory comes from the "No/few smashes while below 50%" standard, along with the note from Short Fuse in Fly's recent post. It could also explain said no/few smashes behavior. (Or do we already know why she does that?)

 

[Fly_Amanita]

Popo's percentage definitely influences her behavior. For example, if both ICs are at 0% and Nana gets grabbed, then if we call the frame she gets grabbed on frame 0, she always breaks out on frame 77. However, if Popo takes a good amount of damage and Nana is still at 0%, she'll break out faster (65 and 56 frames being a couple of the counts I observed right now with Popo at 53% and Nana at 0%).

 

[Gatoray]

I know this isn't the best evidence, but someone messaged me on Reddit with this video they personally recorded while experimenting with this trick. Apparently level 9 Nana powershields lasers:

http://a.pomf.se/bjykye.webm

They further explained that Nana does not do this when set to level 1, but provided no video proof. I realize it's not impossible for level 1 CPUs to powershield, only unlikely. But wouldn't the probability of power shields be evident when comparing Nana levels?

Also, another thought I had. What if this trick is specific to a certain version of Melee? That could explain why some people are unable to really notice anything.

 

[Fly_Amanita]

It's difficult to precisely acquire numbers to support that since Nana apparently doesn't powershield unless she's trying to fight somebody and she could throw out a move right before a laser hits her, denying the possibility of a powershield in the first place, which could make it hard to get data like the approximate percentage of lasers powershielded. If there is a really substantial gap between the percentage of lasers powershielded (e.g. Nana powershields most lasers when she can at level 9 and none when at level 1), a crude test still might be able to suggest that, though. This is something I could also see percentage heavily influencing.

At this point, it'd probably require really substantial data to convince me that this trick actually changes anything at all considering my own experimentation, Nintendude's experiences, and the thoughts of people experienced with memory addresses, like ResidentWaffle and Achilles.

 

[Deleted member]

I know this isn't the best evidence, but someone messaged me on Reddit with this video they personally recorded while experimenting with this trick. Apparently level 9 Nana powershields lasers:

http://a.pomf.se/bjykye.webm

They further explained that Nana does not do this when set to level 1, but provided no video proof. I realize it's not impossible for level 1 CPUs to powershield, only unlikely. But wouldn't the probability of power shields be evident when comparing Nana levels?

Also, another thought I had. What if this trick is specific to a certain version of Melee? That could explain why some people are unable to really notice anything.

Looking over that footage leads me to think that Nana will only powershield when she actually 'turns into' a level 9 computer as other people have suggested. From that and my testing in DEBUG mode today I have some information.

Nana will go through multiple CPU modes - she doesn't stay under the CPU mode NANA all the time.
The CPU mode switch is obviously dependent on the situation, but may also be dependent on Popo/Nana's percentage as others have thought. The CPU modes Nana can switch into are NORMAL*, FREAK, DEFENSIVE, COOPERATE, and STAY. The CPU mode FREAK is when a CPU stays put and attacks opponents if they get close to her. This could explain her not running up to people from time to time and only jabbing them when they get close. The DEFENSIVE CPU mode causes Nana to only use defensive moves such as shields and rolls (I think this is what we see in the video). The COORERATE CPU mode causes Nana to walk around and attack without Smashes. This may be what she does when trying to reconnect to Popo.

Let's say Nana is separated from Popo. The first thing Nana will do is try to find a path back to him. If Nana can not find a path to Popo, she goes into the STAY CPU state (Nana goes into the STAY state and not the NOACT state. NOACT CPUs will not try to recover or DI). If Nana does successfully make it back to Popo, she will enter the NANA CPU state - This is the only CPU state in the game that allows her to perform attacks that the player in her specified port command her to do.

If Nana is separated and the game's code sometimes switchers her CPU mode to NORMAL when Popo or her are at a high percent. When in the CPU mode NORMAL, Nana will start mimicking the moves of a 'normal' computer opponent, reading from whatever level is defined in your port from use of the trick. When either Popo or Nana is at a relatively low percent, The game can switch her to other modes like FREAK or DEFENSIVE as we saw in the video. When Nana is seperated from Popo in the video, she went into the FREAK cpu mode and started jabbing. When she was hit with a laser though, she went into the CPU mode DEFENSIVE.

I encourage you to test this out - Play against a level 9 Ice Climbers and spam projectiles at them (Use a space animal and just laser them). Popo+Nana or Popo alone will powershield a bunch of projectiles, but Nana only will when she is feeling threatened when she is alone. Nana will never powershield lasers when she is following Popo or is in any other cpu mode but NORMAL (When she attacks like every other computer and draws data from a level 9) or DEFENSIVE (when she is using shields or rolls).

*edit: It may have been suspected Nana will go into the CPU mode STRUGGLE (this is when a CPU fights like a normal CPU but is not attracted to fighting humans over other cpus) instead of NORMAL at times, but this is not the case; Nana seems to have a tenancy to target humans when other CPUs are in play.

 

[Encephalon Decay]

Looking over that footage leads me to think that Nana will only powershield when she actually 'turns into' a level 9 computer as other people have suggested. From that and my testing in DEBUG mode today I have some information.

Nana will go through multiple CPU modes - she doesn't stay under the CPU mode NANA all the time.
The CPU mode switch is obviously dependent on the situation, but may also be dependent on Popo/Nana's percentage as others have thought. The CPU modes Nana can switch into are NORMAL*, FREAK, DEFENSIVE, COOPERATE, and STAY. The CPU mode FREAK is when a CPU stays put and attacks opponents if they get close to her. This could explain her not running up to people from time to time and only jabbing them when they get close. The DEFENSIVE CPU mode causes Nana to only use defensive moves such as shields and rolls (I think this is what we see in the video). The COORERATE CPU mode causes Nana to walk around and attack without Smashes. This may be what she does when trying to reconnect to Popo.

Let's say Nana is separated from Popo. The first thing Nana will do is try to find a path back to him. If Nana can not find a path to Popo, she goes into the STAY CPU state (Nana goes into the STAY state and not the NOACT state. NOACT CPUs will not try to recover or DI). If Nana does successfully make it back to Popo, she will enter the NANA CPU state - This is the only CPU state in the game that allows her to perform attacks that the player in her specified port command her to do.

If Nana is separated and the game's code sometimes switchers her CPU mode to NORMAL when Popo or her are at a high percent. When in the CPU mode NORMAL, Nana will start mimicking the moves of a 'normal' computer opponent, reading from whatever level is defined in your port from use of the trick. When either Popo or Nana is at a relatively low percent, The game can switch her to other modes like FREAK or DEFENSIVE as we saw in the video. When Nana is seperated from Popo in the video, she went into the FREAK cpu mode and started jabbing. When she was hit with a laser though, she went into the CPU mode DEFENSIVE.

I encourage you to test this out - Play against a level 9 Ice Climbers and spam projectiles at them (Use a space animal and just laser them). Popo+Nana or Popo alone will powershield a bunch of projectiles, but Nana only will when she is feeling threatened when she is alone. Nana will never powershield lasers when she is following Popo or is in any other cpu mode but NORMAL (When she attacks like every other computer and draws data from a level 9) or DEFENSIVE (when she is using shields or rolls).

*edit: It may have been suspected Nana will go into the CPU mode STRUGGLE (this is when a CPU fights like a normal CPU but is not attracted to fighting humans over other cpus) instead of NORMAL at times, but this is not the case; Nana seems to have a tenancy to target humans when other CPUs are in play.

Hm...Well this is certainly an interesting prospect. Is there any way to incorporate the %-based behavior into this layout as well?

Just another random thought I'm throwing out, but perhaps Nana's solo behaviors could be reminiscent of two computers on the same team?

 

[Syrox]

I totally believe this is real, when Myougi and I were playing, Nana power shielded a knee and also back air edge guarded me another match.