I play League of Legends. Some might claim that I’m obsessed, that I spend too much time playing and thinking about the game. They’re wrong, of course.
In unrelated news, I had this wonderful idea for a post while thinking about League. (The name’s sort of awkward, so let’s go with this.) What does it mean to play League perfectly?
Before starting each game, which lasts roughly half an hour, you pick the champion you’ll play in that game from a list of 159.1 Each champion has a unique set of abilities and a unique way of playing the game, so perfect play must look different for each champion.2
Though the exact details of a champion’s abilities are crucial in the context of the game, they’re not so important for my point here. As an example, let’s consider the features of champions classed as assassins: they’re naturally aggressive and particularly reward players who have mastered the champions and can push them to their limits. Consequently, they’re popular champions to one-trick, that is, to play exclusively.
There are some assassin one-tricks whose play stands above the rest.3 What are the characteristics of their play? What makes them better? Many one-tricks have mastered their champion and can aggressively push them to their limits; one cannot distinguish themselves through slight edges in expressing their champion’s natural identity. Rather, it’s that their play is disciplined, restrained, and controlled—they approach perfection in a way that others don’t.
Perfect play is neither aggressive nor defensive. To call it one or the other is to miss its beauty: perfection never strays from the line that divides aggressive and defensive play. Aggressive when it can, defensive when it must, perfect play simply is what it is. What adjective could do it justice?
Sure this is a nice story, and I hope it seems as intuitively reasonable to you as it does to me. I ought to elaborate on what I mean, as I’m currently only appealing to whatever concept you have of perfect play. Game theory has a precise definition, but I’m looking for something that might be useful to me as a League player, and that thing is probably more broad and less technical.
Let’s start with Connect Four, a simple and hopefully familiar game. The game is so simple that it’s been solved by brute force, and this solution tells us that with perfect play the player who moves first wins. In game theory, perfect play is play that leads to the best possible outcome no matter what the opponent chooses to do, as determined by the solution to the game. For example, in Connect Four the first player has to start by dropping their tile in the middle of the board. Then, no matter the move their opponent responds with, they have a move that guarantees the game remains on a path to their inevitable victory; this is characteristic of perfect play.
Chess is a little more complicated and so it hasn’t been solved.4 This means we don’t know what perfect chess play looks like in this game-theoretic sense. As this narrow game-theoretic conception of perfect play doesn’t apply to a game as simple as chess,5 we’d like to have a broader and more generally useful conception of perfect play.
But first, note that the concept also applies to games without perfect information, such as poker. The key feature of perfect play is its inexploitability: it guarantees the best possible worst outcome, no matter the opponent’s play, or expected outcome in the case of games without perfect information. Indeed, there now exist poker solvers which calculate perfect play.
In search of this broader conception, let’s think about the main weakness of perfect play: its inexploitability, which in turn prevents it from maximally exploiting poor play. For example, in scissors paper rock perfect play randomly chooses between scissors, paper, and rock with equal probability. This will never beat anyone, and it’s certainly not optimal against the strategy of only choosing rock: if your opponent has rocks for brains, the optimal strategy is to stick a paper bag on their head. It’s probably not murder if they don’t have a brain, right? Of course, everything changes the moment they pull out a pair of scissors.
Interesting games aren’t like scissors paper rock: they can’t be so trivially solved, so we don’t have a detailed picture of perfect play. But we tend to develop an intuition for what sorts of play are more and less exploitable by other players at our skill level. And surely the stronger the player the less exploitable their play, which naturally leads to the notion that play converges to perfect play as player skill increases. This notion carries with it the assumption that the space of possible strategies has some sort of topology, some sense of what’s close to what, which you need to define convergence. What this topology might be isn’t clear to me, but this notion of convergence feels so intuitively reasonable that I’m going to stick with it anyway.
In these difficult games, the distinction between optimal and inexploitable play becomes muddy. Good moves in chess are good precisely because they minimise the maximum harm inflicted by the opponent’s best move: this minimax decision rule captures the spirit of perfect play. Indeed, if you give a computer with this rule infinite computing power, it’ll play perfectly.6
Things become more complicated when the margin by which you win matters. In games like League, it doesn’t matter if you win by a sliver or a landslide; you need only maximise the probability that you win. But these complications are beyond the scope of this post, as I only really care about League.
Some maxims, then, to characterise a broader conception of perfect play. Perhaps they’ll be useful.
Don’t be defensive: pressure your opponent, induce mistakes, and exploit those mistakes. Don’t passively allow them to take more than they deserve.
Don’t be aggressive: perhaps your opponent hasn’t made a mistake, but rather has laid a trap. Don’t overextend and give them opportunities they don’t deserve.
Strive to play without adjectives; strive to play inexploitably; strive to play perfectly.
People sometimes like to claim that a beginner could defeat a master by unexpectedly doing something ‘bad’.7 Don’t aspire to such a false and thin mastery: any master who strives towards perfection, upon instinctively disembowelling the beginner, may remark on the foolishness of the claim. But be careful: defeating beginners often engenders poor habits, a misguided pursuit of optimality over inexploitability.
In evaluating play, don’t be distracted by irrelevant details. That professionals make horrendous optimisation mistakes8 is only evidence that these details matter much less than solid fundamentals. As with everything, people at the 95th percentile of skill constantly make basic game-losing mistakes—these are what should worry you. It doesn’t matter that your sword is sharper than your opponent’s when yours lies on the ground and theirs lies buried in your gut.
Lastly, follow the well-trodden path on your road to perfection; the sheer imposing cliffs that border the path are guides, not climbing practice. Use the mistakes of others to avoid being victim to the same pitfalls. Use their wisdom and experience to get unstuck and find your way back to the path.
I’d like to say that these ideas translate naturally to life more generally, but I don’t think that’s quite true. I recently read Surrogation, a book by Suspended Reason, which attempts not to present anything new but to synthesise many known ideas into the frame of surrogation. It’s easiest to introduce the concept through a few quotations.
On games:
Before we begin exploring the surrogation concept, I want to establish the prototypal game in which full-bodied surrogation occurs. This is the strategy game, where each player’s best course of action depends on the course of action other players take. Virtually any interaction between agents can be meaningfully construed as a game of strategy, since agents definitionally possess desires (goals) whose pursuit may conflict or align with other agents’ pursuits, and whose attainment is a product, in part, of those agents’ actions. These games may be as cooperative as trying, on the highway, to avoid a collision, or as adversarial as negotiating a corporate merger. They may be as simple as rock-paper-scissors, or as complex as a world war. Because each player’s desired outcome, and his own best moves toward securing that outcome, depends on the decisions and actions of other agents in the game, he ends up reading fellow players for signs which testify to these decisions and future actions. And because a player’s interpretations of such signs alters his decisions and actions, other players find it worthwhile to actively falsify and manipulate their displays of signs, to alter rivals’ decisions and thereby secure their own goals.
We will focus in part on a specific kind of strategy game, particular to modern interaction, where relative strangers must vet each other during brief windows of mutual exposure, often prior to a high-stakes decision. It is these games in which surrogates are most depended-upon, and most exploitable. Job interviews, romantic dating, and door policies at nightclubs are among the prototypal selection games. This is in contrast to most casual interaction, which may be strategic but is nonetheless typically characterized by agents who know one another well, and are trying to build and maintain long-term relationships. We will build up to full bodied selection slowly, beginning with games involving just a single agent.
On surrogates and surrogation:
In selection games specifically, surrogates can be understood as behavior by the designers, implementers, or enforcers of a game (forming the game’s “incentive structure”). In formal, institutional games, it is the necessary translation of spirit into letter, often accompanied by an amnesia that this translation has taken place, so that the letter is reified as the purpose itself of the game, or the basis itself of success, rather than as a flawed means of tracking and motivating play. But surrogates are also employed, as we will see, in informal selection games—cocktail parties, gallery openings, military battles, children’s games of hide-and-seek—any situation in which players are sizing up and “reading” one another, while producing “writing” for the other to read. And in both formal and informal games, selecting and selected players alike present a front—a public-facing set of selection priorities or qualities which are often only partially related to the real criteria of selectors, or the real qualities of the selected.
Surrogation is about how the images and icons of things come to replace the things they are images or icons of. It is about our removal from private realities, and our reliance on statistical testimonies.
On selection and perception:
The savvy evaluator, rather than subscribing to a naive belief in neutral, uncontested perception, takes an adversarial stance, actively disguising or misrepresenting his own selection procedure, in order to subvert manipulation attempts by evaluated parties.
Alfred Korzybski once said ‘the map is not the territory’: the symbol of the thing is not the thing itself, the surrogate not the thing surrogated. Human interaction is ruled by surrogates and optics because we make decisions based on our perceptions, not reality. We invent metrics to guide and simplify our decisions, and by their very nature these metrics are targets.9 Games are won by targeting metrics, won according to the letter that so poorly captures their spirit. Metrics are obfuscated and letter patched; metrics are identified and new holes in letter found. And so it goes.
Some people like this. Others do not.
Constructed symbolic games, like chess and League, are not like real games: occurring entirely in the world of symbols, the representation of the game state is the game state itself. The spirit is the letter, and so they sidestep the problem of surrogation entirely. Moreover, they have what C. Thi Nguyen calls value clarity: the goal is clear, and so players need only concern themselves with how they achieve the goal within the strictures of the game.
These types of games are very popular. I think this is precisely because they offer an escape from the complexities of the real world and real games. There’s no need to reckon with the complexities of surrogation, no need to even work out what to value: the goal is merely to win, and to win one must merely play well.
This is not to say that these games are easy. Rather, they’re easy to practise. Within the limited domains of these games, we may attempt to approach perfection; outside of them, the concept of perfection hardly makes sense. Perfect according to what values? Perfect according to whom? In these games, we can often look to a ranking system that quantifies player skill. Though rank is a surrogate for skill, it’s a tight surrogate10—it might not be quite so tight a surrogate as visual perception is for the world, but it’s close enough. And this soothing clarity can’t be matched by real games, making such games increasingly appealing in an increasingly complicated world.
In a way, the practising of these games is practice for the practising of real games. They can teach us to take things seriously, to analyse our decisions and fix our mistakes, to approach everything with a plan and intent, to control our emotions and play with discipline, and many things besides.
But constructed games are not real games. They’re an escape from real games, no matter how much I rationalise their utility. If we don’t care for doing real things, this isn’t a problem. If we do, if we want to do real things, we have to play real games. Yet reversed stupidity is not intelligence—something isn’t wrong just because it’s a rationalisation. So I’ll keep playing League, keep attempting to approach perfection, and hopefully learn some things in the process. As long as I improve, it’ll serve as an existence proof that deliberate focused improvement is possible. At the very least, I’ll learn that it probably won’t teach me anything.11
An ever-increasing number!
Indeed it’s obvious that perfect play at each moment must depend on the entire history of the game up until that moment, or at least the part of the history that can be known to the player, as the game does not have perfect information.
For the League players out there, I’m thinking of Zed99’s Zed, Beifeng’s Qiyana, and Pzzang’s Yasuo. And I’m quite uncritically accepting the narratives Midbeast presents for each of these players. Coach Curtis claims—I think quite reasonably—that you need to be a very strong player, roughly Master tier, the top 0.1% of dedicated players, to understand the intricacies of such high-level play. I’m certainly not that good at the game!
Though we’ve solved all configurations of seven pieces, so we do know what perfect play looks like in this context—there’s a lot of slow shuffling of pieces.
One could claim that chess is complicated, but one would be wrong. Chess has perfect information and the space of possible moves is extremely constrained. This simplicity is why computers find the game so tractable.
Of course, convergence remains an issue. There’s no guarantee that its play ‘smoothly’ approaches perfect play as you dial up the computing power—it might at some point realise that an entirely different style of play is better than its current style.
Google isn’t particularly helpful here, but I recall this claim in the context of a beginner swordsman—hence ‘disembowel’.
For the League players out there, I’m thinking of this game between RNG and HLE in Worlds 2021, when Deft built Infinity Edge third after building Galeforce and Wit’s End. Then he built a Guardian Angel fourth. For the non-League players, Deft forgot that his choice of second item, Wit’s End, meant that Infinity Edge wouldn’t work properly as a third item, and then didn’t build a fourth item that would fix the issue. Less appalling optimisation mistakes occur more regularly in professional play.
This resembles Goodhart’s law—‘when a measure becomes a target, it ceases to be a good measure’. And the moment a measure plays a role in decision-making it becomes a metric, a target measure: it ceases to become a good measure.
For example, in the process of learning new skills your ranking may dip in the short term as these new skills occupy your attention and impinge on your ability to demonstrate known skills.
This seems unlikely. There’s a lot of solid educational League content out there with some pretty general morals on improvement, including Coach Curtis and his League podcast.
In particular, the most recent episode has an interesting section on alignment. Developing alignment in how a team sees the game is crucial for turning that team into coordinated and excellent one, and it’s a task that takes months and years of cooperative discussion on things like champion identities and roles, and win conditions.
From Surrogation on alignment:
Alignment itself—too often reduced to merely an economic or robotics problem—is one of the difficult and profitable problems to solve. Alignment is what enables cooperation between agents, and therefore is foundational to the creation of functioning societies, institutions, and superorganisms. And what I think is becoming clear, with our notion selection games, is that representation undergirds alignment in non-evolved systems. Where evolution tests the real, intelligence infers from the apparent; in between lies a world of difference.
Alignment is not just an issue in artificial intelligence. It’s an issue everywhere, and we don’t have any easy solutions.