Kal
Smash Champion
- Joined
- Dec 21, 2004
- Messages
- 2,974
Introduction
I find this issue to be something many people have an egregious misunderstanding of. In particular, most players seem to think reading an opponent is equivalent to just guessing (or informed guessing) and committing 100% to that guess. I call this model the "Discrete" model for reacting to your opponent: "I think my opponent will do X, so I will do Y." I will discuss the "Probabilistic" model, which generalizes the Deterministic one. In particular, I will explain why the Probabilistic model is superior to the Deterministic one, provide theoretical examples of how to apply the Probabilistic model, and the provide a methodology for applying it to actual gameplay.
The Probabilistic Model
The Probabilistic model works as follows: you simply assign a probability of an event occurring, and you assign a value to your possible reactions. It's not as complicated as it sounds. To show this, I will apply it to a simple gambling game.
Let's say I have a fair, two-sided coin. If I flip heads, you have to give me $10. If I flip tails, I have to give you $20. So then your decision making is simply whether you play the game. To calculate whether you should, you simply evaluate your odds multiplied by your expectation, keeping in mind that you assign a negative value to losses:
(.5)($20) + (.5)(-$10)
and the resultant expected winnings is $5. In mathematics, we would call this your Expected Value, and we would be a bit more precise about what all of this means. But, for practicality, it's easy to think of it this way: as you play more and more games, the amount you win will approach an average of $5 per game.
So, in fact, my "Probabilistic model" is just an application of first semester probability. Hopefully I'm not accused of plagiarizing Fermat or Pascal.
The Discrete vs. the Probabilistic Model
As a general rule, guessing what your opponent will do is probabilistic. Unless your opponent is exceptionally bad, you will never, with absolute certainty, know what he will do. So you will necessarily apply probabilities to it. "Reading your opponent" is not a matter of guessing what action he will take in a particular situation. Instead, it's about recognizing patterns (e.g. that your opponent techs in place frequently, or that he rolls after every third fair). This isn't to suggest that you cannot guess what your opponent can do accurately. Rather, it's to suggest that people are simply not so predictable. In fact, what most people actually do is simply observe what their opponent does most often and assume that this is what they will do. This is necessarily a bad strategy for reading your opponent; even if what you guess they are going to do is their most likely course of action, the payout from your reaction may not outweigh the cost for being incorrect.
In other words, the Discrete model fails to factor in risk of failure, because it implicitly assumes with 100% certainty that your guess is correct. More importantly, any application of the Discrete model exists within the Probabilistic model. "My opponent will do X" is equivalent to "The probability that my opponent will do X is 1, or 100%." In this sense, the Probabilistic model generalizes the Discrete one, and so, unless the Probabilistic model is hopelessly impractical, it is clearly the superior model.
Theoretical Applications
Luckily, it is not so impractical, and these theoretical applications should make that clear. For simplicity's sake, I will keep the tree of possible outcomes short, but this can be made as complex as necessary to help determine which solutions are optimal. We'll assign kills a value of 1 and deaths a value of -1.
Let's suppose that you are Sheik, and that you dthrow your opponent. Your opponent has a tendency to tech in place. In fact, he techs in place 70% of the time. Conversely, he techs left 15% of the time and he techs right 15% of the time.
Suppose further that you have three options: up smash in place, dash attack left, and dash attack right. Finally suppose that, if any of these attacks hits, you will get a kill, and if any of these attacks whiff, you will die. Then your expected payout for up smashing in place would be:
(.7)(1) + (.15)(-1) + (.15)(-1) = 0.4
Similarly, your expected value for dash attacking left would be:
(.7)(-1) + (.15)(1) + (.15)(-1) = -0.7
And the in the same vein, the expected value for dash attacking right would be -0.7. In other words, your best option (of these three; again, this is a very theoretical example) is to up smash in place, because its payoff is greatest.
However, this is not a particularly realistic example. We can come up with a more realistic, still theoretical example. Suppose you are Falco, and you are standing behind a Fox in his shield. You assess the situation (because you have infinite time in this theory-crafted example), and decide that Fox will bair 75% of the time, usmash 20% of the time, and spot dodge like a noob 5% of the time. You decide that, if your opponent bairs, a roll behind him would successfully gain you a kill, but a grab would get you killed; if your opponent usmashes, a grab would sucessfully get you a kill, but a roll would instead get you killed. Finally, you realize that, if your opponent spot dodges, a grab will get you nothing, but a roll will get you killed.
While this decision tree is noticeably more complicated, the following table might make it clearer:
roll|+1|-1|-1
grab|-1|+1|0
And so we simply apply the probabilities, as in the above example. For rolling, your payout is:
(.75)(1) + (.20)(-1) + (.05)(-1) = 0.5.
On the other hand, for grabbing, your payout is
(.75)(-1) + (.20)(1) + (.05)(0) = -0.55.
In other words, in this (again, simplified) example, rolling is clearly your best option.
In-Game Applications
While it's more or less impossible to really quantify all of this information, this decision-making model is still quite practical. In general, you simply want to maximize your gains and minimize your risks. One example of a clear application of this is whenever you have a guaranteed option instead of one which requires you to make a guess. For example, while Marth does have combos against Fox and Falco which utilize throws other than uthrow, your payout from uthrow is virtually guaranteed (determined more by whether you mess up than what your opponent does). Comparatively, with throws like fthrow and dthrow, you are instead dependent on correctly guessing what your opponent does. Since the payouts would be equal (we're expecting a kill in the end, since Marth's combos on the space animals very often lead to death), it's clear that Marth would prefer to uthrow.
However, many players would find this application too obvious. Of course you would go for a guaranteed thing over something which requires guess work! Let's consider a less obvious application (I will use Marth again because I am not terribly familiar with the rest of the cast): suppose you grab Sheik at 0% damage. The obvious response is to uthrow. From this, you can utilt. If the utilt hits, you are guaranteed a combo of at least 60% damage. However, she can nair out of this and hit you before your utilt comes out, in which case you will surely get grabbed and suffer a similar combo. On the other hand, if you shield after the uthrow and she does not nair, the fight will reset and Sheik will only receive the damage from the uthrow. If she does nair, then you will get a regrab and from here you can uthrow, and the utilt to follow is guaranteed to hit.
So, we could possibly assign estimates of the damage received and given as payouts for this, but that goes against the idea of quickly and practically assessing these sorts of situations. Instead, simply consider that, should you go for the utilt, you will suffer a combo every time your opponent decides to nair out of shield. Instead, if you simply shield, and your opponent does nair, you will successfully combo them by regrabbing. But, if your opponent does not nair, the gameplay simply resets, and you lose nothing. Thus, the risk is clearly minimized by the latter option, with the reward being roughly equal.
Now, should you notice a pattern in your opponent's gameplay then you can adjust your strategy. If, for example, your opponent does not seem to ever nair out of the uthrow, then you can simply always go for the utilt. Conversely, if your opponent seems to always nair out of the uthrow, then you simply always shield.
Conclusion
When applying the probabilistic model, a player needs to avoid being what is called "results oriented." It's pretty much exactly what it sounds like. If the option you are choosing to take pays off in the long run, then you should not become upset with yourself if it does not pay off a particular time. That's not how probability, and particularly reading players, works. And, in the same vein, you can't assume that just because something worked once that it was the right thing to do. In other words, the logic "it worked didn't it" and "that didn't work" should not influence decision making within this model.
I hope people find this guide helpful. It should hopefully shed some light on what is the best way to make use of your ability to read your opponent and recognize patterns. Rather than simply taking the course of action that will solve what your opponent will most likely do (which, as mentioned before, is a hopelessly terrible methodology for reacting to your opponent), it should allow for players to maximize their rate of success.