Theory

[icycalm]

Here's a question for you: do you think it is possible to figure out what kind of game someone is playing, just by looking at his hand movements on the controller?

[Vert1]

Sure. If I saw someone holding up on the analog stick with the A button held down on a Dreamcast controller I'd suspect the person was playing Daytona USA. Racing games are probably the easiest genre to recognize through hand movement.

Although there are things that look easy to identify that are not. I've played games like RE4 and Soul Calibur 2 in ways attributed to other games. I played RE4 "Mario Party 8 style" (each person holds half the controller) with a friend and I've played SC2 with bongo drums.

My friend one time muted the tv (I assume your thought experiment would require this) and we all noticed how intense the button clicking was. We were playing Smash Brothers Melee. So, fighting games would probably be an easy genre to identify with all the directional inputs that make the thumb roll in a certain way.

edit: someone mention the Konami code.

[EightEyes]

Given a competent player who is playing the game seriously and using a genre-appropriate input device, I think you'd be able to narrow it down to a fairly specific game style.

Since we define styles of games by "what the player does" in the game, it stands to reason that we ought to be able to discern the style of game by watching what the player is doing from moment to moment. The graphical context shouldn't be necessary. If it is, and we can't deduce game style from control input, then that might be a good reason to question some of our genre distinctions.

[Archonus]

Ugh, this thread reminds me of how ridiculous most actors look when they pretend to play video games on T.V. and movies. I once saw a character mashing the X button of a PS2 controller while playing a racing game. I shit you not.

[icycalm]

Given a competent player who is playing the game seriously and using a genre-appropriate input device, I think you'd be able to narrow it down to a fairly specific game style.

Since we define styles of games by "what the player does" in the game, it stands to reason that we ought to be able to discern the style of game by watching what the player is doing from moment to moment. The graphical context shouldn't be necessary. If it is, and we can't deduce game style from control input, then that might be a good reason to question some of our genre distinctions.

I couldn't have put it better myself. My theory is that is should be possible to develop a computer program, which, when fed with the output file of any controller operated by a competent player, would tell you exactly what kind of game the player is playing.

Indeed, you could go the other way around. I.e. you start by compiling a huge database of such output files, each file corresponding to different games and different players, and then the computer should be able to sort all files that resemble each other into categories, each one of which would represent a different genre or subgenre.

This would work for all kinds of games, with no exceptions, as long as you observe the two rules that EightEyes mentioned: competent player and appropriate input device for each kind of game.

This approach, if done properly, would be so sensitive that it would be able to distinguish, not merely between fighters and shooters, for example, but between shooters that are vertical and manic, horizontal and manic, horizontal memorizers, goofballs like Ikaruga and Exelica, or between old-school SF-style fighters and modern combo-heavy ones, etc.

In reality, trying to discern between different genres and sub-genres can sometimes get tricky, but with a program like that entire new horizons of analysis would be opened up. I need to write more about this.

[To The East]

It would certainly be possible to do this. Here is a link to a paper describing how to use acoustics as a keylogger:

http://www.cs.berkeley.edu/~zf/papers/k ... -ccs05.pdf

It doesn't seem like it matters who is typing, because it is only measuring the relative differences in key presses. So, I don't think you would need a database of different games and players. The difference between controllers might be a problem, because a PS3 Dualshock's buttons are in vastly different positions than a Arcade Pro's buttons. However, it wouldn't be a problem to tell the difference between the various kinds of Dualshocks even if they made one out of steel and one out of rubber.

[Afterburn]

Given a competent player who is playing the game seriously and using a genre-appropriate input device, I think you'd be able to narrow it down to a fairly specific game style.

Since we define styles of games by "what the player does" in the game, it stands to reason that we ought to be able to discern the style of game by watching what the player is doing from moment to moment. The graphical context shouldn't be necessary. If it is, and we can't deduce game style from control input, then that might be a good reason to question some of our genre distinctions.

Can we take this idea to its logical conclusion and claim that all video game genres can be reduced to certain patterns of button presses and stick movements? That genres are just collections of particular controller inputs? I.e. to play a shooter game is simply to press buttons "this way" and to play a racing game is to press buttons "that way" and so on.

[icycalm]

Yes, we have already said that, but not exactly in those words.

I have not yet drawn all the conclusions that can be drawn from this. I am getting there, though.

[Vert1]

Outside of certain genres, I think it would become almost impossible to name what kind of games people are playing. An area we haven't explored yet are genres that could be mistaken for other genres being played. I'm surprised no one has mentioned that video game developers are given a control scheme they need to conform to (primary action button, secondary action, etc.).

How would you tell someone is playing an STG if they are doing a run where they don't shoot (i.e. bullet eater mode in Ikaruga)? The game inputs might look like someone who is playing a game like Super Monkey Ball. Nobody has even mentioned how increased difficulty will most likely require players to speed up (altering) their directional inputs which could be interpreted as belonging to another genre.

[icycalm]

Outside of certain genres, I think it would become impossible to name what kind of games people are playing.

No, you are wrong. RTSes will all have similar input patterns. Turn-based strategy games likewise. Racing games likewise. I cannot think of a single exception to this rule. And if the genre itself may be too broad in terms of input patterns to categorize, the SUB-GENRES will not be. In the end, the program itself will tell you what the genres are. If your previously accepted genres do not match up with the computer's, you will simply adopt the computer's scheme.

I'm surprised no one has mentioned that video game developers are given a control scheme they need to conform to (primary action button, secondary action, etc.).

I don't understand what you are saying here. Sid Meier is given "a control scheme he needs to conform to?" Or Tsuneki Ikeda? No dude. They just do whatever the hell they want to. Trust me on this. Besides, even if they weren't, this would have no effect on the present discussion. You can't design an RTS that has input patterns that look like a racing game, no matter how much someone might be pressuring you to do so.

How would you tell someone is playing an STG if they are doing a run where they don't shoot (i.e. bullet eater mode in Ikaruga)? The game inputs might look like someone who is playing a game like Super Monkey Ball.

First off, your example is a poor one. Bullet eater Ikaruga input patterns would look nothing like Super Monkey Ball's. Try to visualize them and you will see.

But even if it turns out that Bullet eater Ikaruga patterns look like those of some, say puzzle game, that would only shows us the kinship that exists between those two games (Bullet eater Ikaruga being viewed as a different game to normal Ikaruga -- which it is. Bullet eater Ikaruga has an additional rule that says "You can't shoot", which makes it a different game to Ikaruga.) And this is what I mean when I say that this approach would open up new horizons for game criticism and analysis.

Nobody has even mentioned how increased difficulty will most likely require players to speed up (altering) their directional inputs which could be interpreted as belonging to another genre.

I don't see why speeding up would alter directional inputs. Speeding up is speeding up. In any case, we are talking about a huge database here. The end result would be a series of graphs (looking much like sound wave graphs) averaged over many different players, difficulty levels, etc.

It's really hard to explain what I mean without actually carrying out a demonstration and plotting some test graphs. Without that, you need to be able to visualize the thing for yourself. I can clearly see the input patterns transformed into graphs right now, in my mind, but I can't communicate this process to anyone else...

[ganheddo]

Yes. The more elaborate and distinct the input device and input patterns, the easier and more accurately you could differentiate certain kinds of games. E.g. one-switch games would be the most difficult to differentiate, yet by analyzing the timed sequences, it should be possible to at least tell if it's an adventure, strategy, or action game. And while I could imagine extreme cases where differentiation would be impossible, these would be just highly unlikely exceptions from the norm.

That it is perceived as being impossible or very hard, just shows how abstract and standardized our input devices are, especially compared to output devices. It's obviously easy to tell the kind of game someone's playing by taking a look at its output. This will probably vanish when devices allow more distinct input patterns (e.g. it's pretty easy to tell which one of the wii sports games someone's playing).

PS: How about telling a game by looking at a player's facial expressions or just his eye movements?

[bullethell]

Yes. I would easily recognize a MMO gamer looking at the way he uses his keyboard. Pressing F buttons multiple times every ten seconds? Definitely a MMO gamer doing his daily grinding duties.

Furthermore, I would be able to describe his type (player attitude) by examining his keyboard inputs for like an hour or so. If he never typed a single word during that period he must be the solo player (he neglects other people in a virtual world) and chances are that he is going to quit the game really soon.

[Muzozavr]

The mouse is going to be the hardest to identify, because it depends on movements that look too similar to each other. However, on anything with more different moves (e.g. keyboard) you could tell, I think.

[icycalm]

You could "tell" with anything. It's only a matter of the scientists doing their job properly. This is a simple enough matter that they should be able to handle it. The very fact that a mouse is being used narrows down the genre considerably. Not to mention that you could stipulate that ALL kinds of games be played alternatively with ALL kinds of controllers, and then have the computer sort out al the data for you and make cross-sample matches between different kinds of controllers.

All the above is elementary for those who know statistics. All objections in this thread come from people who don't. The study is very very feasible, and if extensive will give amazingly accurate results.

[Muzozavr]

I agree that mouse movements for different genres would be very different. An exhaustive study would be accurate. But can you percieve that difference with your naked eyes? A computer could tell, yes. What about your eyes?
On a keyboard, and on a controller, the movements and button presses are easier to see by far.

The very fact that a mouse is being used narrows down the genre considerably.

Could you provide me an example of a genre that can't work with mouse being your primary way to control. A mouse-controlled platformer WAS made BTW and it was good. (Search for "Maverick flash game", it's quite hard)

[icycalm]

But can you percieve that difference with your naked eyes? A computer could tell, yes. What about your eyes?

WHO GIVES A FLYING FUCK? A COMPUTER WILL BE DOING THE REGISTERING ANYWAY JESUS FUCKING CHRIST.

Could you provide me an example of a genre that can't work with mouse being your primary way to control.

I am not going to "provide" you with anything. I am not here to take retards by the hand and explain to them how the world works. If you can't figure out at least ten examples off the top of your head, you should stop posting in this forum.

Either way, stop posting in this thread because your stupid shit is getting on my fucking nerves.

Fuck.

[icycalm]

More insight from the same guy today:

It is sometimes difficult to measure but 2-ALLing Ketsui is definitely much, much, much harder than clearing Galuda or Futari Maniac, for example. But yeah, I do think someone who can beat the hardest game thus has the ability to beat all the easier ones, if they use similar skills (for example Ikaruga uses more methodical skills whereas Caves require more danmaku skills and that's different).

http://www.cave-stg.com/forum/index.php ... 58#msg9158

[bunuelo]

The kind of (computer) analysis you are talking about is roughly what I do for a living, although with data that are not of the type discussed here. In practice, the most basic machine learning algorithms typically solve this type of problem with ease, even if the data is largely noisy or absent.

There are at least two approaches to the classification/clustering application you are describing. The difference between them is whether or not they are supervised. That is to say, do we have labeled instances of each "class" that exists in our model of the world (in this case, each genre or sub-genre)? If so, we are doing classification. If not, we are doing clustering.

The primary benefits of having labeled instances are (1) it makes (automated) learning much easier and (2) it is significantly easier to test whether our tool is doing a decent job (empirical validation). The downsides are (1) the time and expertise spent annotating data instances and (2) the requirement that we know a priori what classes can exist. Disadvantage 2 is particularly nasty; we must restrict ourselves to some finite set of classes that we can come up with, and the classes must be good at discriminating between data instances.

Unsupervised learning (clustering in this case) has its own advantages and disadvantages, but for brevity you can just imagine the previous paragraph entirely negated.

Two additional disadvantages of unsupervised methods are their immaturity in the literature and that their results can be harder to interpret. In particular, an unsupervised method can tell you that it thinks games 123,455,... should be grouped together but it cannot tell you to label that group "side scrollers" -- you would have to do that on your own. (If this sounds easy, consider the difficulty in getting a group of people to agree whether a given game belongs to a given genre.)

Assuming supervised learning, my experience leads me to expect that you can discriminate between genres easily and with high accuracy. Probably just having counts of each button press (with each direction or half-quadrant of a directional input, and discretized or weighted values for held buttons) would be sufficient, even if the names of the buttons were not provided for any instance. Including simple dynamics (the sequence of events without their actual timing) typically increases accuracy if necessary.

In any event, all of this stuff is indeed basic and has been done for other (but similar) types of data for several years now.

[Pilgrim]

Bunuelo, couldn't you essentially scan the inputs once unsupervised and then, using those chunks that appear, be able to create groups knowing what games fall into what category. Once we have the unsupervised data it would be much easier to decide where to draw our genre lines for the supervised data. For instance we could place Super Mario Brothers as an instance of side-scroller platforming no matter how the data was chunked. Once we pick the titles we're sure fit the right genre, we would be able to intricately draw lines between genres that the unsupervised chunking missed. Basically, we would be able to formulate the classes for the supervised data with a greater knowledge of what distinctions we need to apply to those classes based on how the data was chunked unsupervised.

[bunuelo]

Most good unsupervised methods "believe" that the number of clusters is infinite, and that only by seeing more data instances can previously-unseen clusters be discovered. Your suggestion restricts the number of clusters in our (supervised) model to the number of clusters present in the initial "seed" instances.

Moreover, there will invariably (in my experience) be clusters that are difficult to name. While "side-scrollers" may not be one such cluster, "puzzle games" may very well be -- in fact, what an expert calls "puzzle games" will most likely include several clusters, each of which is hard to name. It is not unusual to have a cluster whose most succinct description, by an expert, is simply a list of its members.

For these reasons, your suggestion is actually the worst of both worlds. It suffers from clustering's inability to know what a given cluster should be called and classification's inability to know about classes it has not yet seen.

A more attractive middle ground is active learning, which is sufficiently explained by its Wikipedia entry.

[icycalm]

Moreover, there will invariably (in my experience) be clusters that are difficult to name. While "side-scrollers" may not be one such cluster, "puzzle games" may very well be -- in fact, what an expert calls "puzzle games" will most likely include several clusters, each of which is hard to name.

This is because "puzzle" is not a videogame genre. It is quite a complicated issue to explain, and it presupposes knowledge of videogames which no one other than me currently has, but I will get around to explaining it quite soon, and all the relevant issues that go with it.

[icycalm]

http://postback.geedorah.com/foros/view ... 889#p12889

I can't help you with your memories of me. I am not dense -- I am the most intelligent person you will ever interact with. That's all I have to say on the subject. The only thing people who call me dense, or any other names, accomplish is to lower my estimation of them since they are not smart enough to see the obvious. In your case, however, you can say whatever the hell you like. I have long ago pigeonholed you and I don't see any reason to ever change my estimation, especially given the limited nature of our interaction.

As for your question, it depends on what you mean by "substantial". Within the genre of shooting games, a change of perspective means a change of subgenre. That seems like a pretty "substantial" change to me. Viewed from the perspective of "action games" on the other hand, the change is indeed rather minimal. But we were not discussing things from that perspective in this thread. It's the same with something like Death Smiles and Espgaluda, etc. Even though mechanically many horis and many verts may appear, on the whole, extremely similar, they FEEL extremely different (again the "extremely" applies WITHIN the genre, outside of it you need to drop that word) -- and it is the "FEEL" that, when all is said and done, determines a game's genre -- as I will demonstrate at length in an upcoming essay probably titled "Pressing Buttons", or something to that effect. But you really should have caught on to what I am saying by now -- at least if you are not as dense as you've been acting in this thread.

tl;dr version: I won whether you like it or not, so suck it up, dude.

And to be even more specific, read this thread:

http://forum.insomnia.ac/viewtopic.php?t=2528

So basically, to determine whether it is justified to call Pulsar Viewpoint's spiritual successor we would have to conduct tests to determine to what extent a typical Viewpoint input file is dependent more on the perspective or THE REST OF THE MECHANICS. If we find that it is the perspective (which is what I think we will find, as should be pretty obvious), then the MEAT of the game, its ESSENCE, is the perspective, hence a game WITHOUT this perspective CANNOT be called a spiritual successor, since it betrays the original game's SPIRIT.

And I repeat, you should be able to figure this out intuitively by playing both games and being an experienced gamer -- this, after all, is what reviewing is all about -- but in any case one day we'll get a computer printout that says "icycalm was right", and that will be the end of that.

You can't read the connecting discussion if Recap hasn't given you access to that forum, but I still want to save these replies of mine here for future reference.

[Showa]

Moreover, there will invariably (in my experience) be clusters that are difficult to name. While "side-scrollers" may not be one such cluster, "puzzle games" may very well be -- in fact, what an expert calls "puzzle games" will most likely include several clusters, each of which is hard to name.

This is because "puzzle" is not a videogame genre.

Traditional videogame genres have only the vaguest of relations with clusters of games that are played with similar player inputs, and expecting insightful automated creation of named, abstract clusters is silly: success of this clustering experiment would be being able to tell that game A is objectively more similar to game B than to game C, and describe what inputs are different.

Apart from the speculations about discovering similar games in different genres, there's a lot of difference between the details of apparently similar games that people who think in terms of "puzzle game" and "side scroller" cannot suspect.
For example, several famous Cave shmups share a fairly similar 2-button main weapon system (press briefly for a wide shot, hold for a narrow beam, hold another button for wide shot autofire), but ESPGaluda has a button to switch modes (which happens quite often) while DoDonPachi DOJ has a button for the traditional sort of bombs that are only used in dire need: presumably, the former is going to be pressed more often and at more variable times than the latter.

A concrete instance, like comparing ESPGaluda with DoDonPachi DOJ shows a number of theoretical and practical issues with clustering games by player inputs:

• Not all playthroughs are similar. In an arcade-style shmup, dodging in a slightly different way and getting hit causes a local disturbance (respawn invulnerability, not shooting for a while, etc.) and long term consequences when the game ends earlier. Similarly, hitting with one more bullet can mean the difference between killing an enemy or not, which can often cascade into not getting a powerup, not getting enough points to earn an extra life (which is almost equivalent to dying), being rated bad and sent to different stages (e.g. Psyvariar), and so on.
  So, what is noise and what is important? How can we aggregate input traces that have little in common to find the patterns of a "typical" game? Do we need to look at game state?
• Not all players play in the same way. For example, movement patterns in a shmup aren't very constrained, and sometimes other features like the use and abuse of mode switching in ESPGaluda tends to be left to the player's responsibilities; experienced players are likely to converge to a close approximation of the optimal way to play a shmup for completion and score, but there are many ways to get there and not all players are talented, experienced, and competitive. Someone might even misunderstand the game and handicap himself (e.g. not knowing how the "guard barrier" in ESPGaluda is activated).
  How can the clustering be objective? Should we base the comparisons on the input of a representative player? What player? What does it mean to study a game, if it is something different to every player? Unlike objective elements of a game like rules and graphics, player inputs are strictly subjective.
• Given the extreme variability of whole playthroughs, considering smaller units of input might be more productive; but how can they be discerned without examining game state? For example, if I keep shooting for several seconds with frequent short left and right adjustments around a constant center position, it might be twitch dodging between thick bullets, or tracking a small moving target, or a way to spread my bullets a little wider, or something else. Is it a relevant distinction? Game inputs are easy to measure, but they need game outputs, and probably the human touch of analysis, to be meaningful.