By Katherine Morayati

Mention “interactive fiction” to someone of a certain age – too young for Infocom, too old for whatever they’re calling “millennial” today — and you’d be forgiven a little skepticism. You might know 1995 as the year hobbyist IF sprouted from near-barren cultural ground, but they might know 1995 as a particularly dubious time in interactive entertainment. A lot of things were converging: the mass-mainstreaming of the World Wide Web, the sudden feasibility of sophisticated graphics and video on everyday personal computers, the now-quaint lack of irony regarding anything futuristic. In short, interactivity was the new hotness, and everyone wanted it. Historical sites could be “explored” via CD-ROM recreations that invariably resembled Doom levels. Musicians yoked all sorts of PC-based frippery to their albums. (Among the auteurs who bit: Prince, David Bowie, Laurie Anderson and Peter Gabriel.) Huge amounts of effort – much crass, some genuinely good – spent on what would become obscure, if sometimes fascinating, ephemera. (Let it be a cautionary tale for those repeating history.)

But the Holy Grail was, and perhaps is, interactive film. Video games, after all, have aspired to cinema since the technology advanced enough to make such aspirations feasible; arguably, modern blockbuster cinema aspires in the other direction. Each field tends to think the other is easier than it is. But the leap from choose-your-own-adventure books to film – even short film – is particularly difficult. What can be conveyed in writing with a couple dozen words and imagination serving as cast, crew and designer is exponentially more difficult (and time-consuming, and expensive) in film. And all this assumes a single-player experience – how does this translate to a movie theater without devolving into Twitch Plays-style bickering?

Enter Interfilm, which signed a deal with Sony to outfit more than 40 theaters with a proprietary joystick rig that allowed audiences to vote for where they wanted the film playing to go. The idea wasn’t exactly novel. Interactive theater had been done; 1985’s The Mystery of Edwin Drood, based on an unfinished Dickens novel, allowed the audience to vote for the desired murderer, romantic pairings and secret identities, oddly reminiscent of the same year’s adaptation of Clue. (For an IF-world take on this, see Dietrich Squinkifier’s Coffee! A Misunderstanding.) Interactive film, too, had been attempted as early as 1967, with the Czechoslovak World’s Fair entry Kinoautomat. All these were comparatively well-received. So what went wrong? Intra-corporate bickering, perhaps; a failure to recall history, littered with short-lived “immersive” gimmicks like Emergo, Percepto, Illusion-O, Smell-O-Vision…; or most fatally, a level of pandering to the 18-34 male demographic resulting in such noteworthy cinema as robot grossout-revenge flick Mr. Payback, which Roger Ebert (perhaps foreshadowing his later stance on games?) called the worst film of 1995. It’s a movie with the rare distinction of being so bad, it’s literally unwatchable; when your film relies on proprietary, brick-and-mortar robo-sadism installations, that’s not gonna translate to YouTube even if someone wanted it to. And its counterparts, Adam West vehicle Ride For Your Life and blow-shit-up excuse Bombmeister (stymied by a certain incident in Oklahoma City), have also disappeared, save for their trailers.

The exception is 1992’s I’m Your Man, released on DVD during yet another renaissance of the idea, Tender Loving Care and Point of View. (The idea seems silly, but so did reviving 3D, right?) The film can still be found used, for those with morbid curiosity, questionable senses of humor, friends with both, and no aversion to the seller slipping up and sending a Leonard Cohen film instead. It so happens that I am all of these things. (If you are not, the intro can be watched online.)

The most striking thing about I’m Your Man is, despite its tropiness, the utter lack of cynicism on the creators’ part – especially given Hollywood’s habit of prospecting every new technology for potential revenue streams and places to stake fan properties. The making-of video is stunningly earnest; everyone except the composer, who cheerfully admits his mediocre work-for-hire, everyone talks about the film in the breathless, uncontained voice of the true believer. In interviews, the creators admitted to fantasizing about satellite versions. Someone who probably is Bob Bejan – because who would impersonate someone on something this niche? – went so far to post an Amazon review of the DVD: “…As stupid as it looks, we spent A LOT of time thinking about it. To be totally honest, a bunch of us who were there are still thinking about it. … There is little question that we were ahead of the curve.” The subject: “Interactive narrative is HARD.” The score: 4 stars.

“Ahead of the curve” may be a little generous. The plot of I’m Your Man is thin, the standard love quadrilateral between the villain, the MacGuffin, the woman and the flirt. Jack, a hapless would-be audience stand-in whose main personality trait is being bad at flirting and, thanks to Kevin Seal’s VJ background, evoking a wry Carson Daly; Leslie, a transplanted Elaine Benes type and the actual audience stand-in; and aforementioned Richard, who is played by The Master from Buffy. There is action of sorts, but it’s either seconds-long gags or questionably filmed, somewhat appropriative fight scenes. There is urbanity, via a self-consciously pretentious gallery opening; there is humor, of the dubious one-liner sort. There is sexual tension, of a sort, but it’s mostly goofy and unsuccessful skirt-chasing or the kind of self-serious seduction (“Let’s dance… while we talk”) that’s all but disappeared from pop culture. It’s played for laughs, but with juuust enough plausible deniability that it is intended to be serious.

What, then (besides snark) would compel someone to watch this thing? Well, interactive narrative is indeed hard, and a lesson is always useful.

Branching

An axiom of interactive fiction: branching paths exponentially increase the amount of writing one has to do, not to mention the number of variables and plot points and world state the writer and simulation has to keep up with. Now imagine all that, with shooting. Little surprise, then, that I’m Your Man branches in name only. The entire structure fits on the back cover of the DVD and is a standard friendly gauntlet. Most of the primary plot events are fixed; the only choice is whether to watch them or leave them to exposition. What actual plot choices exist rejoin the main storyline relatively quickly. The effect is something of a Rashomon narrative: the story happens, and your main decision is who to follow through it.

Critics did not like this. (“This experience is not like watching a real movie … it is more like rooting for a basketball team,” wrote Caryn James of The New York Times.) Yet there’s something to the idea; filmmakers might be limited by shooting costs and logistics, but writers have no such limitation. You can get more plot and characterization mileage from 60 or 120 minutes than you can from 20. With this sort of branching you can avoid two choice-structure pitfalls at once: the railroading of the traditional gauntlet, the uncontained sprawl of the time cave. And you can get a lot more mileage if your assumed audience is not the kind of 13-year-old proto-bro who’d be found these days in the audience of Suicide Squad. Imagine something like Photopia rearranged chronologically, but with a choice of whose story to follow. Or perhaps something like Exhibition drawn out into a full-length story. (Both these examples are parser-based, but if anything this would be an even better fit for choice, less constrained by an object model.)

Consequence

A side effect of the aforementioned bro-focus (brocus?) is that every narrative tends toward wish fulfillment. Consequences are irrelevant. Failure is abhorrent. Tragedies don’t happen. Every choice must somehow demonstrate how virile and awesome the PC, and by extension you, are. I’m Your Man follows this model. Most choices result in victories of some sort, however improbable: sudden feats of pepper spray, out-of-nowhere ninja moves. No matter how little game Jack has – a would-be flirtation scene finds him boasting about how, ah, fast he can get stuff done — Leslie can eventually acquiesce – it’s harder, and fits the story less, to make her not. There is no “bad ending,” no failure state; the player chooses which character achieves a “victory,” executed in-story via deus ex machina twist. The implications for CYOA writers are obvious.

Immersion

Most negative reviews of interactive film have a curious, near-phallic fixation on the joysticks used, making what I’m sure seemed the obvious leap to video games, and their apparent inherent dudeliness. Technology has advanced considerably in the 20+ years since, but it’s still not quite possible to make the decision process physically seamless. There will always be an inescapable physical sign of gimmickry at work.

Interactive films addressed this by making whatever implements exist part of the story. (I’m Your Man does it via campy tutorial-style intro.) But perhaps it’s best to do the opposite. Put another way: a compelling story goes much farther than a villain who breaks the fourth wall to inform the audience that, if they don’t make their choice fast, their seats will dissolve into acid. Or, as it were, their computer.

By Lisa Brunette

For the past eight years, I’ve worked as a narrative designer in the segment of the digital games industry referred to as “casual gaming.” My audiences have typically been either families (for console games), young women and girls (handheld console games), or women over 40 (downloadable and app games). In this article, I’ll concentrate on the evolution of HOPAs, short for hidden-object puzzle adventure games. This genre was popularized from 2005 to the present by the dominant publisher, Big Fish, which in its PC origins also had a player audience primarily made up of women over 40, with their specific preferences and needs. A lot of work went into getting developers, with their typical “gamer” appetites, to transition away from combat and into intrigue, with story and gameplay focused on solving mysteries and puzzles.

In the wider scheme of game evolution, the HOPA can trace its roots to early text-based adventure games. But today’s typical HOPA player would likely shy away from the fantasy settings that characterize most early examples, such as Dungeon before the advent of personal computers and Zork afterward. In the evolutionary branch I’m analyzing here, supernatural mystery or mainstream adventure in the style of Indiana Jones have been the prevailing genres. But while the adventure served as a sort of frame or ancillary, the hidden-object portion of the HOPA came first.

In a hidden-object puzzle, players search a complex visual scene to spot items on a given list. The first games were little more than digital variations on “I Spy” games the player audience would have already been familiar with, as they were popular mainstays in print newspapers and magazines throughout the U.S.

But even in this nascent incarnation, the impetus to add story was strong. Early developers started out with an idea for scavenger hunt-style gameplay. But they wanted to include a mystery element as well, as a narrative bolstering the game. While these developers were working from what had inspired them and what they wanted to see in a game, data supports this creative move. We know that story acts as a powerful motivator for players to both purchase and play a game. Today we can point to the Entertainment Software Association statistic that an “interesting story or premise” is among the top reasons players buy a game, and this is across the industry— all players, all genres.

In the HOPAs of 2008, the relationship between text and visuals was one I’d describe as separate, and not entirely equal. It paralleled the relationship between the story and the game play: The story was text-based, and the gameplay was a visual experience, except for a few lines of instruction text. But even in the play, the visuals were mainly static, with few animations. This story-through-text approach was the norm for a few years, with the typical in-game journal serving as the vehicle for story even as more animations became possible. In the below example from Mystery Case Files: Ravenhearst, the story is delivered via the character’s diary, relating her backstory and impressions. Players find the diary entries as they progress through the game, learning more about the character and events.

As the HOPA evolved, the game play took on some of the characteristics of console role-playing games, with movement and exploration within and between scenes now possible. Thus the true adventure portion of the HOPA came to the fore. Players could find an object in the opening scene, store it in their inventory, and use it in a later scene. For example, in a typical HOPA quest chain, players pick up a handkerchief in scene one and realize they can use it to clean a dirty mirror they come across three rooms deeper into the world. However, player discomfort with too much of an open world demanded that the games remain linear. Hence the rise of the “door puzzle,” a mini-game that unlocks a room in a house, as one common use, giving developers a gameplay-based way to control players’ navigation through the world.

The story aspect had evolved as well, toward better overall integration with the gameplay so that it wasn’t relegated to the journal. Players were instead treated to full cinematic cut scenes, not just as intros and outros but within the game, triggered by player action and furthering the story and goal. At this stage, the relationship between text and visuals was one I would describe as complementary, as in the example below. Here the character Death in Riddles of Fate: Wild Hunt speaks directly to the player, who is now fully immersed in the world as a player character.

We also raised the quality bar on the premises and plots, putting more resources into the narrative design team. When Big Fish recruited me to third-party production in 2011, I was the only narrative designer. In just two years’ time, I was managing a staff of four.

At this point, my team was charged with an interesting task: To reduce the amount of text in all our HOPAs. There were many compelling reasons for this charge. First was efficiency; the games were clocking in at anywhere between 50,000 to 120,000 words—about the length of an average novel—for only seven to twelve hours of play. The company’s promise is a “new game every day,” and it took a tremendous amount of time and resources to write, edit, and test basically a novel’s worth of text for each game, plus the costs of translating into multiple languages and then re-editing and testing.

But more importantly for the craft of narrative design, we recognized that ours was primarily a visual medium. While all of us were essentially writers drawn to our work because we loved words, we also knew that visual storytelling held more sway in the context of a game over mere words on the screen. Players had been telling us for years that they did not want to read a lot of on-screen text, or that they ignored the text in the journal. It was a bit of a paradox, as they sometimes said they didn’t care about the story, but this usually meant they were equating story with text. We knew that story could make or break a game, and that the more visual the story, and the more it was integrated with the play, the better the player experience.

Just one example that showcases the results of all this work—and exemplifying the modern HOPA—is ERS Games’ Puppet Show: The Price of Immortality. You can download and play the demo for free here, or you can watch the walkthrough here.

Getting back to that idea of the relationship between text and visuals, the end result is, in the best games in the genre today, a well-woven tapestry of cut scenes and game play. And that old-fashioned journal? We found we could remove it altogether.

By Hugo Labrande

>solve zork

Recent advances in artificial intelligence allow us to solve problems and design technologies that not so long ago were science fiction. We ask selfishly: what does it mean for parser IF?

Deep learning

Throughout most of the history of computing, to get a computer to solve a problem for you, you first had to think about it really hard, then program the computer to perform some steps in the right order. This worked very well — it pushed humans to come up with creative concepts, do some theoretical work, and ultimately advance science.

But think about it: is this how you were raised? Are our children programmed, perfectly obedient and only as intelligent as we are? Do we even understand what is in their brains? This is the paradigm shift brought about by artificial intelligence: create structures to hold information and a great number of parameters (basically, a brain!), and a way to have these parameters adapt locally and internally as a function of external feedback.

This is the idea behind neural networks, proposed in the 1980s: neurons hold a small quantity of information and interact with each other, reconfiguring pathways and weight given to each element to get better. The main problem with neural networks is the processing power involved in making a computer that’s at least decent in accomplishing its task: you have to provide millions of examples, with enough diversity that the neural network will know how to deal with the full range of situations. Then for each of these examples, you have to actually perform the simulation of each neuron’s behavior and their interactions and compute the adjustment in the neurons as a function of the result. This limits neural networks’ efficiency, and they were abandoned by the ‘90s as simpler methods gave results that were just as good. Enter deep learning, and the neural networks make their comeback in style.

Now might be a good moment for a caveat: I am very far from being an expert in these matters, but I was lucky enough to be able to attend a conference by Yann LeCun, one of deep learning’s best researchers and head of AI at Facebook Research, and the whole thing blew my mind! A lot of things were explained simply, and there were so many examples, including some that inspired this article. So I’ll try to explain this as best as I can, and refer you to the slides if you want to know more.

The “deep” in “deep learning” stands for “a large number of processing layers”. The idea is to have a lot of specialized (and trainable) components that perform specific operations, such as filtering, on part of the input; higher-up layers aggregate results from these layers, which makes them able to look at higher-level features of the input. Just like in the visual cortex: simple cells in the eye get a few pixels of the image, more complex cells aggregate these small pieces of information to get features of an image, then more complex cells recognize some features as part of an object, etc! This visual analogy is actually precisely why deep learning has been so successful in image recognition and classification, achieving incredibly low error rates.

Deep learning is very recent — no more than five or ten years old. It still requires a lot of computational power, but several racks of current graphics cards are enough to train efficient silicon brains. The results are astonishing: deep learning is already a reality, hiding behind Siri and Google Now, automatic picture tagging on Facebook, and the victory of AlphaGo against Lee Sedol; in the future, they’ll be in self-driving cars, chatbots, and MRI machines.

Building a computer with common sense

 

Interestingly, deep learning methods are also applied to natural language problems, which are one of the most difficult classes of problems that we may want to teach computers how to solve. This includes problems that human brains solve daily: parsing sentences, linking words to abstract concepts, performing small logical deductions, perform abstract reasoning, and even translation! This is the heart of human communication, reasoning, and consciousness, much more than playing Go or identifying a picture of a banana; for this reason, Alan Turing proposed that the ability to chat with a human so well that the human doesn’t know if she is speaking to a computer (the famous “Turing test”) is the definitive sign that computers are as smart as humans.

Hence, giving a computer “common sense” is still hard and unsolved. How do you teach a computer to know that if Jack has a glass of milk and went to the office, the glass of milk is also in the office? How do you teach it that if you close the oven, the food is still in the oven? That water does not burn? That people cry when they are sad?

We’d need a way to show the system thousands of situations that highlight properties of the physical world; thousands of different objects and settings, so that it learns what usually goes with what; thousands of occurrences of objects interacting with each other. And maybe give it a way to interact with such objects, so that it experiments and gets better at learning physical properties in an unsupervised way, just as AlphaGo got better by playing against itself thousands of times. But if you drop them in the real world, it’s going to take a lot of time and a lot of effort (as well as a lot of broken things): you have to learn how to operate arms, vision, etc. A 3D simulation, maybe? It’s been done: researchers have used the Unreal Engine to teach a computer how physics worked, so that the computer ended up predicting that an object was going to keep falling until it hits the ground. But this requires image processing, even real-time processing abilities.
But why don’t we make the computer play parser games? This is learning through play, using the textual nature of parser games to teach the system natural language, and common sense reasoning through puzzles; and there are literally thousands of games on the IF Archive, so lots of material to learn from!

This is not a totally original idea. Myke, a user on intfiction.org, recently opened an intfiction.org thread to ask for help with his research, which consists of exactly that: using parser IF to teach an AI about language understanding and physical properties. As he notes, “the dream scenario would be to let the system train on a set of IF games and then test it on games that it has never seen before and see it perform well.” (We wish him all the luck in his research and hope he’ll show us his results!)

Furthermore, as he notes, there are least two recent papers that have looked at text adventures as a way to train an AI to understand text and physical properties. However, these works train a neural network to get better and better at a game: the resulting neural network does not learn how to play games in general. The first one, by MIT’s Karthik Narasimhan, Tejas D Kulkarni and Regina Barzilay, attempts to teach a system how to play a simple parser game (called a MUD in the paper) in a house (sample quests: you are hungry but not sleepy, and there is a bed in the bedroom and an apple in the kitchen), then a larger game in a fantasy setting; the system eventually manages to complete the game on each try, once it has completed the game 50 times. The paper attempts to see if any of the gained knowledge could be “transferred” to other games, but their method is baby steps, as they only show that if you move the bedroom and the kitchen (but keep the same objects and descriptions), the system learns how to solve that game faster. The second, by Ji He, Jianshu Chen, Xiaodong He, Jianfeng Gao, Lihong Li, Li Deng and Mari Ostendorf, uses choice/hypertext games where the choices are randomized (to prevent the machine from simply remembering the number of the options); the computer then gets better and better, and seems to also gain an understanding of the situations, as the second part of their experiment (replace actions with paraphrased descriptions of the actions and see if the system is still able to get a high score) shows. All this is preliminary work, but rather exciting nonetheless!

What and how a computer would learn from parser IF

Let’s step back a little bit and break down what this all might mean, from a text adventure author/aficionado perspective. In particular, we’ll take a closer look at different aspects of text adventures and player behaviors, and see how they might apply to an AI able to play parser games. This is also helpful to take apart mechanisms of parser IF and see what they demand (and how hard they can be) for a completely uninitiated user.

Input

First of all, we should ask: is this even doable? The only way to know for sure is to attempt to build it. Deep learning is an interesting and powerful technique, but there are a lot of challenges. For instance, the syntax of commands: normally these consist of a verb followed by an object, but some verbs have prepositions (e.g. ask X about Y, put X on Y), which complicates matters because the English syntax is not the same for every verb. And for that matter, doesn’t this imply that the system should first figure out what’s a verb? Or even what’s a valid word?

From various attempts at coding AIs that can play video games, the process might go like this: a completely newborn system would probably ignore the command line for a long while, then figure out that typing is an interesting idea, then type absolute garbage for a long while until it gets the idea of typing a command starting with “x ”, then go on an examine binge until it finds a word that matches, hopefully learning that words that are in the text are more likely to get responses. But how does it figure out “take” when most games don’t even mention it? An interesting alternative would be to bootstrap this process by enforcing that the system can only type a word from a selection of words – for instance, all the standard verbs, plus any word seen in the output of the game. Then the system would “just” have to figure out that >swim never produces anything useful, as opposed to >examine or >north.

Even with just that, we are talking hundreds of verbs, hundreds of nouns and countless words from the game’s output, while previous attempts to make computers play video games focused on console games with half a dozen choices at each frame, instead of tens of thousands. The large search space simply means that more computations would be needed to figure it out. But in theory the computer will figure it out eventually, given enough time, processing power, and probably memory – a non-trivial task.

Reinforcement

But let’s carry on and assume we have a system that knows it should type words. We then have the question of the “objective function,” of the feedback that lets the system modify itself: what kind of things should the system pursue, and what things would let it know it’s on the right track? First of all, it seems like we should restrict ourselves to playing games with at least one preferable end state, or our poor machine will get confused trying to figure out what it needs to do to finish the game. Exit Aisle and Galatea (wouldn’t a computer talking to Galatea be the most meta thing, though?) However, it would be very interesting to train a system fully, then make it play these games: will the system loop infinitely, looking for more endings, or will it think “yup, that one, that one is the best ending” and stop? Which will it choose?

Back in the day, most games had a scoring system that would reward you with points for quests, subgoals, secrets, etc. This kind of feedback seems like a natural candidate for our AI: the higher your score, the closer you are to the end of the game, and maximizing your score means you completely beat the game. Furthermore, perhaps timed sections will teach the AI that time should not be wasted and the number of turns should be kept to a minimum. But eventually, the system will have to play modern games, perhaps with puzzles but no automatic score reward; how would the system react to this kind of game? Obviously it’ll continue playing, but it will probably walk around aimlessly and unfocused for a while. Then it could figure out other, more subtle encouraging signs: managing to take an object, reaching a previously unreachable area, a NPC that says “thank you” — at which point the AI would be starting to think like a player!

On a side note, one cannot talk about things that signal that you’re on the right or wrong track without talking about the cruelty scale. Merciful and Polite games seem to be the best for this system, as once it has figured out that nothing can be done when dead except restore or undo, it’ll probably have learned that dying is bad. Tough games seem a bit more complicated: the game will give warnings before the act, but if the system doesn’t register them, it will perform the action, then keep on playing and remain locked out of victory for thousands of turns. Nasty and Cruel are even worse, but again, they are a challenge even for human players! The mere problem of detecting that you’re stuck seems like a very hard problem: how do you know that the game will never let you win, how do you know that there’s not a small non-obvious action that you have forgotten to do? This is often solved by glancing at a walkthrough – which obviously is not applicable in our case. So this particular problem seems to be particularly prickly.

Objects

One thing we have not mentioned yet is the relationship with objects, which is crucial in parser games. We humans generally acquire object permanence, i.e. the mental process which says that objects have an existence even if they are not currently seen, before we are two years old — rephrased differently, this is a skill that takes humans two years to learn! Because traditional parser games are primarily interested in object manipulation, this seems like the perfect way to teach the concept to a machine, and this is probably useful in a lot of applications which deal with giving the machine something that resembles common sense.

This will presumably mean that, for a long time, the system may go to the room where the troll guarding the bridge is asking for a fish, look for a fish in the room, go to the next room, see a fish, and fail to connect these pieces of information as the troll is not in view anymore. On a more basic level, the concept of inventory is one that should be very hard to grasp for the machine! I mean, think about it: there’s the object you need, but you typed “take” and now the object is gone. This is clearly not good, right? But eventually, the game should figure out that typing inventory will make the object appear – and even better, that if you take an object in a room and then move, the object is still in your inventory.

This should be a major breakthrough for the system, and besides, it’s necessary to solve puzzles (or even figuring out that taking objects is a good thing and not a bad thing!). No doubt that learning this skill will be a large advancement in giving a computer common sense.

Puzzles

It is now time for our system to attempt to solve puzzles. Much has been said about the design of puzzles in adventure games elsewhere, and it is interesting to think about the different pieces of advice that have been given and recast them in our context of an AI learning how to play parser games.

For now, we just wish to highlight a few things on fairness. Puzzles are often qualified as “fair” or “unfair”; what would happen if an AI stumbled on an unfair puzzle? A puzzle can be unfair for a lot of different reasons and may stump our system in numerous ways. The puzzles where you are not warned at all before dying, or have to die and retry, may be solved easily by the system (once the command “restore” is known), as a computer would presumably not get as frustrated as a real player would. The puzzles where no hints are given are a bit harder to solve, but ultimately, if the machine is smart enough not to loop and try the same cycle of actions over and over again, it may well stumble on the right solution given enough time. The same applies to puzzles that require external knowledge or rely on in-jokes: the “hint” may as well not exist. One of the trickiest types of puzzles is the puzzle that requires a very specific formulation, a specific verb, or even those for which a perfectly sensible solution should work but doesn’t. These puzzles are notoriously frustrating for humans, who eventually turn to a walkthrough, and the question of how a machine would solve it is interesting. It may require the machine to learn synonyms of a verb or of a noun (which is a very abstract and very high-level concept), and enough stubbornness to keep going down a path while metaphorically screaming “this should work!” Hence — as it is for humans — it is probably best to avoid unfair puzzles.

Before talking about fair puzzles, let’s deal with the question of mazes, as these are inevitable when playing a large number of parser games. The standard maze is the one that only needs to be mapped; this implies that the AI system figures out what mapping is (unless the strategy of drawing a map has already been found just to be able to move around) and has enough memory to draw one. We then increase the difficulty by having non-reciprocating room exits; then perhaps some random events, exits or rooms, which increase the difficulty quite a bit. Then, imagine the system’s confusion when encountering “a maze of twisty little passages all alike”: it has to realize that these are indeed different rooms (and not just the same room), and find a way to distinguish them, rediscovering the standard “drop objects” technique: this strikes me as quite a feat of logical reasoning. Then there is the great number of mazes that require the player to realize that this is not a standard maze and that there is a trick. These require even more logical deduction, and presumably great maze experience, for an AI to be able to figure this out. So mazes themselves seem to be a very complex topic, one whose resolution would be reasonably considered to be an advancement of science – can you imagine having an AI crafted specially to solve mazes for you?

Let’s assume we stick to fair puzzles, then. Discounting clues given to the player, one can theorize that fair puzzles are the ones that rely on common knowledge or common sense, and that make sense once solved. This is a prime example, and maybe the main reason, why parser games are an amazing possibility to teach AI systems common sense: by learning through play and experimentation, the system will notice patterns that correspond exactly to common sense, and will learn and perhaps reinforce some new things, sometimes by blind luck. Things may start slow, by learning how to solve “measure 5L of water with 2 cups of capacity…” puzzles a thousand times, but presumably also by learning how to retrieve a key with a bobby pin and a newspaper. Things can then progress to more advanced concepts, like light puzzles, or the use of mirrors: games which rely heavily on simulationist aspects, like the Zork games, can be great to learn about concepts like light sources, liquids, burnable objects, or even ropes, as they allow a lot of experimentation and deal with fundamentals of our physics. The caveat is that the game has to have very few bugs, so that the AI is not misled into thinking that these bugs also happen in reality. Increasing the difficulty, one could think of associations of such concepts, as in the solution to the maze in Photopia; and finally, more complex cultural concepts (setting off the fire alarm makes people vacate the premises). Ultimately, there are so many games, and so many puzzles, that the sum of the total common sense contained in the IF Archive is probably enough to make a really smart AI!

Implementation

We wrap this up by talking about the problem of game implementation. What happens when the game is full of bugs, typos, is finicky about grammar or formulations, or didn’t implement any of the obvious solutions for a puzzle? Or just when the game is just not very thoroughly implemented? The human player then thinks “this game is buggy, this sucks — but I should be more stubborn and more careful” – i.e. maybe drop ideas less quickly than with a game which is otherwise superbly implemented. If this kind of “meta-gaming” is not done, much frustration can occur: you can search blindly in the dark for a long time, only to find out that the thing you tried in the first place was actually the right solution, but needed a minor adjustment to work!

Can an AI system reach this level of awareness? This reasoning is very meta, relying on the human aspect of game-making: “this game is not reality, it is something that a human has made, and I know that planning every single answer is not possible (or: this human was not very thorough), so I’ll suspend my disbelief and keep playing but keep it in the back of my mind to advance”. Two things can then happen. The first one is that an AI never realizes this, and keeps being tripped up by this: it drops solutions too quickly and never tries them again, doesn’t manage to solve a puzzle in a game because it learned in another game that this solution doesn’t work, etc. Differences in implementation create differences in the simulated game, and from the point of view of the AI, inconsistencies in the common sense; it would have to be robust enough to recover from these, which is very hard to imagine as it is supposed to incorporate feedback to its very core.
The second possibility would be that the system is indeed able to perform this logical jump, and realize that these are only simulations imperfectly created by humans (whose names are given in the first paragraph of the game). Presumably, the AI would adapt its play in function of what it perceives the style or quality of the game to be, or even the author; but fundamentally, this possibility means we were able to create an AI that has now realized it has been living in simulations this whole time. Science fiction then dictates a bad ending for humans – so please, for the love of god, polish your games!

Bloom: Enduring Experience in Episodic Dynamic Fiction
By Cat Manning

Caelyn Sandel’s Bloom, an episodic semi-autobiographical piece of dynamic fiction about her gender transition, plays with the experience of occupied empathy. It is the story of a specific protagonist, Cordy, and even readers who have experienced dysphoria cannot have the same precise embodied experience that Cordy does. By envisioning Bloom as a series, rather than a single game, and releasing episodes periodically (every couple of months), Sandel creates an experience for her readers in which Cordy’s experience must be grappled with; it cannot be digested in a sitting and then compartmentalized. The game uses first-person pronouns — as opposed to the second-person narration commonplace in IF — to move the reader through Cordy’s experiences rather than effectively ceding them to the person beyond the screen. In so doing, Sandel reminds players that Cordy is her own person with her own agency, and that while a player can experience her story, it fundamentally is not theirs to control.

The episodic nature of Bloom requires readers to sit with painful experiences that Cordy experiences — for original readers, to sit with them for months — and the story refuses to resolve them neatly or quickly. In the early episodes, a coworker, Dane, makes transphobic remarks the player can later cite as one of the incidents that tipped Cordy off that she might be trans. In a later episode, Dane tries to confront Cordy about it, but a player is only given the option to refuse to talk to him, in increasingly creative ways. Later, at the very end, Dane apologizes for his behavior, and due to the episodic nature of the series, Dane’s apology felt more genuine to me, rather than a plot point; I can see him educating himself or trying to apologize to Cordy before this final conversation.

There is an option that suggests Cordy’s internal monologue during Dane’s initial confrontation; it is crossed out. Later on, certain options remain greyed out, inaccessible to Cordy based on how much she trusts Dane and their work relationship. Where Cordy was originally powerless to prevent Dane’s comments, it is now the reader who is powerless to change her experiences. Readers might have access to how she feels, but they cannot force her to say something that would compromise her safety or her sense of self. A great deal of Bloom, in fact, deals with what the player can and can’t control. The introductory epigraphs are unskimmable; they appear, phrase by phrase, with time delays. At several moments in Bloom, the player can try to “make” Cordy take a particular action, only to be told several times it’s not possible for her; if pushed too far, she snaps at the player — a reminder that while you have agency over the story’s pacing, you don’t have agency over the story itself; the reader becomes a character in the story, one who doesn’t have the right to force her to take an action she’s uncomfortable with. Whether a player is cis or trans, they are not inhabiting Cordy’s body, with all the specific, personal difficulties the story lays out; a player can close the browser window at the end of the session.

The fact that Bloom is an episodic, rather than a chaptered piece, contributes further to this.  The piece is not only segmented into narrative sections but partitioned in ways that required original readers to wait for the next piece, to inhabit Cordy’s discomfort over a period of time and space. If a reader has followed Bloom since its inception, the wait between pieces creates a kind of space in which Cordy’s story takes the time to breathe. Sandel refuses to provide readers with a narrative that can be resolved quickly, easily, or in one sitting. It’s a particular perspective on the personal, one in which the reader must linger in a highly individual narrative for days or weeks at a time and share their own emotional space with Cordy and her author. Even the end of the current series ends on a note of possibility and futurity rather than a neat conclusion.

[Disclosure: Cat Manning supports Caelyn Sandel on Patreon.]

The IF world, like most artistic fields, is seasonal, and as in music and (to an extent) film, August is a slower month, full of what David Rakoff called “the opposite of hanging out.” If fall and spring are full of new content, awards and the occasional conference or two, late summer is that in-between season, one that looks languid on the surface but conceals a lot of hard work. Dozens of authors, as you read this, are preparing competition entries for the September deadline, or solidifying commercial pitches, or — for those really ahead of schedule — getting their work playtested.

If you’re like me, you’re taking a lot of breaks from being hard at work for such edifying pursuits as playing Minesweeper ripoffs and looking at online auctions for swing coats. But if you’re not like me, you’re using that time to read Issue 64 of SPAG — one I’m especially proud of!

For Issue 64, we’re taking an especially broad view of interactive fiction and its connections, both obvious and not, to other fields. This issue features the dubious, beyond-spotty history of interactive film, the evolution of storytelling in hidden-object games, and the applications of parser games to artificial intelligence research. Of course, we’ve got plenty of more traditional coverage as well, including a Specifics entry on Caelyn Sandel’s episodic piece Bloom and an interview with Brendan Patrick Hennessy, whose Birdland flew away with an entire gaggle of XYZZY Awards, as well as other, less forcedly metaphorical praise.

After you’re done reading, perhaps you’d like to contribute to our next issue? Issue 65, like this one, has no formal theme (as we’ve seen, these things tend to come together organically), but as always, welcome are:

• SPAG Specifics on stories of your choice. These are less traditional reviews and more in-depth critical pieces on how a particular piece does what it does.
• Interviews and/or reviews of figures in the IF world and/or adjacent to it, defined broadly.
• Live coverage, if you live in an area with a significant live interactive fiction presence. This can range from exhibitions to conference coverage to performances to whatever the world dreams up. (Free pitch idea: if you’re a reader in the Toronto area attending the 2016 Wordplay Festival in early November who is not me, I’d like to hear from you.)
• Essays of any kind. The more unexpected, the better.
• Basically anything you can think of related to interactive fiction will be considered!

As always, I welcome pitches by and about women, people of color, LGBT and otherwise underrepresented writers. Also: there is payment commensurate with standard online writing rates.

Send pitches to spag.mag.if@gmail.com. There’s no deadline, but I’d love to hear from you! In keeping with our rough quarterly schedule, Issue 65 will likely arrive around late fall or winter. (What this means for you: anything related to 2016’s competition entries is probably best suited to #66.)

Thanks for reading as ever! We hope you enjoy this issue, and send us the makings of another great one.

(Top Threes is a recurring interview feature in which we ask authors and other members of the interactive fiction community to talk about their favorite things, in their work and others’.)

Brendan Patrick Hennessy’s Birdland, his second competition entry, swept the 2015 XYZZY Awards with a near-record-setting six awards (second to Matt Wigdahl’s Aotearoa in 2009), including Best Game, Best Writing and Best PC/NPC. (Inspiring the following totally trivial but also totally adorable stat: 2015 is the first year the PC and NPC awards went to both halves of an in-story couple.) Since its release the game’s inspired an official epilogue, near-universal glowing praise and countless fan-works.

In keeping with our tradition of interviewing top XYZZY and competition finishers, we asked Hennessy to share his top three…

SPAG: …works of IF that inspired you?  

BPH: 1. Bee by Emily Short. I love it. It’s so wonderful. My favourite from Short’s entire oeuvre probably. It was the first time in the IF world I’d ever really encountered such a well-realized coming-of-age story with such a clearly defined protagonist, and it just totally blew me away. There’s so much about how it’s structured that really resonated with me and that I ended up trying to incorporate into my own work. The way it’s so clearly broken down into scenes, the mechanical cycle that propels the game forward, the sheer size of it. It’s just so good and I really hope a properly working version of it reappears one day.

2. Skulljhabit by Porpentine. It’s actually kinda hard to pick out just one Porpentine game, because pretty much every game of hers did something to expand my ideas of what IF in general and Twine specifically could be. But there was something about Skulljhabit. Something that really grabbed me, some vibe that I often catch myself trying to replicate in some small way, and I’m still not 100% sure what it actually was. Was it the setting? The tone? The opacity of the underlying systems? The deft use of randomization? Perhaps it was just the phrase “Correctly written by the Skull Commissioner”, which will be lodged inside my head until I die.

3. 80 Days is so damn good it makes me weep and I’ve played it like seven times. The fact that something so massive and so engrossing was created by actual living human beings is a source of constant inspiration for me. “A handful of regular nice people created this masterpiece just by working hard and being good,” my brain reminds me. “What’s your excuse?” And I’m like, “uh”

…of your favorite/least favorite YA books of the “truckload” you read while researching Birdland, as you shared earlier?

1. The Miseducation of Cameron Post by Emily M. Danforth. This is the one I completely fell in love with. It’s an astoundingly beautiful novel about growing up gay in rural Montana in the 1990s. And it goes places. Whatever heights you imagine are beyond the reach of YA, this book hits them.

2. A Love Story Starring My Dead Best Friend by Emily Horner. So the classic queer story cliche, right, is to just kill a character as some kind of tragic punctuation mark. But this book takes death as a starting point and turns it around into a story that’s really precious and sweet. It’s about moving on from loss in an unfair world and that’s an idea that’s much more compelling than “HEY YOUR LIFE IS MISERY AND EVERYTHING IS SAD FOR YOU.”

3. I’m not going to throw shade on any specific least favourites, but I will say that all the worst ones I read made the same cardinal mistake of trying to literally impersonate teenage speech patterns. This is completely impossible and no adult will ever accomplish it.

…scenes or material left out of the final Birdland cut?

1. The campfire scene was originally going to have even more endings than the dozen or so that are in the final game. This included an ending where you explicitly came out of the closet to Bell, and one where you actually kissed her at that point in the game. I think in the scene they both would have been really sweet moments but from a broader story perspective it was still just too soon to give Bridget that moment. The other thing was that those branches were going to be behind tough statistics checks, but it was really jarring to have that mechanic intrude on that scene. It’s supposed to be a quiet moment where you’re letting your guard down, and inserting “ERROR YOU ARE NOT TENACIOUS ENOUGH” just completely kills that vibe entirely.

2. I cut an entire day from the game for pacing reasons. Most of the daytime stuff still wasn’t finished but I did have the seventh dream planned out. It was going to be an Indiana Jones style thing. To this day I mourn the loss of the phrase “The purpose of the archaeologist is to arrest the rise of fascism in Europe?”

3. The original ending was on the last day of camp and it had a great moment where all the counselors are back to normal and they’re all confused that camp is over so soon. Like, “Oh, is it just me or did camp feel exactly five days shorter than normal this year?” “Weird, I was going to say it felt four days shorter.” There was also a nice little goodbye with Mackenzie in there where she tells you to follow her on Instagram because she posts “weird pictures of clouds and normal pictures of weird clouds.”

…scenes or moments in Birdland you’re most proud of?

1. There’s a branch in the beach scene where if your melancholy is high enough you can actually break down and cry a little bit about your whole situation. I like how that branch plays out in general, but my favourite thing about it is that it’s a reward that the player has to earn. Like, the mechanic in this scene is: If you raise your stats and play the game well enough you unlock the ability to cry. I don’t know — there’s something I really like about that as an approach to having emotions in a game.

2. I was very happy with how the detective dream turned out because secretly I viewed it as a kind of do-over for Bell Park, Youth Detective. It actually has a much more complex structure than that original game, and as a result it’s a lot more fun to play around in. Weirdly I think it’s also more satisfying as a mystery story, even though it follows dream logic and doesn’t have a “real” solution.

3. That college student dream has so many amazing moments. The bird plan being derailed by hormones, Bridget’s huge out-of-nowhere rant about what it feels like to have a crush on someone, the gibberish poem that you recite in English class. And possibly my favourite little detail of all, the dream logic stage direction: “(You finish your beer and toss the bottle over your shoulder. It flies off into space.)”

…TOTALLY OFFICIAL pieces of Birdland canon?

Officially I decline to answer this question because all canon is illusory. But unofficially here is a 100% canon summary of Liz’s post-Birdland experience:

2015 LIZ: Huh, why do I get such an overwhelming feeling of joy and excitement when I think of my lesbian friends in a relationship? Guess it’s because I’m a really good ally and absolutely no other reason besides that.
2018 LIZ: oh.

…works that missed out on XYZZY nominations you think voters should check out?

1. Fabricationist DeWit Remakes the World by Jedediah Berry. This one slipped under a lot of radars I think since it came out in the middle of IF Comp, but it was a phenomenal piece of sci-fi with lots of interesting and evocative worldbuilding, and some really effective use of music/styling. I loved it. (See also Kevin Snow’s Beneath Floes for a really good Use of Multimedia that didn’t get its due.)

2. Scarlet Sails by Felicity Banks. I actually almost didn’t include this one in here because I was just assuming it had been nominated. But somehow it just missed out, which is a shame. The setting is rich and the adventure is rollicking and the whole thing just captures the feeling of perpetually on the edge of horrible piratey disaster.

3. The ClickVentures. All of them. Clickhole is one of the most consistently funny sites online, and their CYOAs are bizarre and brilliant. My favourite one is probably Can You Survive Seeing Grease on Broadway? (Which is technically a 2016 release, but whatever.)

Also, since we’re on the subject of comedy I’m going to give a bonus 3a shoutout to Bill Belichick Offseason Simulator. Jon Bois is too damn funny for this earth.

…things you want to write IF about someday?

1. 4th-century Chinese history. You ever read about the Sixteen Kingdoms? That shit is unreal.

2. ALL EMOJI GAME. 💯🔣📲

3. Long-term I would kind of like to maybe just go full Degrassi and write some kind of big sweeping seasons-long Canadian teen soap opera. I think that would be a lot of fun. I think that would be the complete and utter end of me as a human being.

…pieces of advice for IF authors?

1. aaaaaa

2. aaaaAAAAAAAA

3. AAAAAAAAAAAAAAAAAAAA

[EDITOR’S NOTE: The author may be being a bit modest here. Further (excellent) advice can be found here, for those interested. Also, aaaaaaaaaaa.]