Electric Dreams, Part 3: Alien Ideas For Player Expectations

Whether or not it’s taken over the industry yet, artificial intelligence and other experimental ideas have been on the mind of people in games lately. So far in Electric Dreams we’ve focused on why it’s so hard to get innovative and risky new ideas into games, but some games seem to manage to push the limits further than others. We’re going to look at a couple of games trying to do this, how they manage player perception, and talk about a new kind of game development that might help risky ideas find their ways into games.

A few years ago I found myself at a London games event talking to someone from Creative Assembly. They had a new project, an incredibly secret new project, that they were all very excited about. All they would tell me was that it involved some kind of creature, and that they had worked so hard on the AI for it that people invited to play would spend long periods in a single room, fascinated by this animal, trying to understand how it behaved and how they could exploit it. It was the game that was to become Alien: Isolation, and even long before it was announced everyone at Creative Assembly knew that this game was selling one thing above all else: intelligence.

Alien: Isolation is the most high-profile game of late to sell itself on the back of ‘cutting-edge AI’. There’s a lot of misconceptions about the game and just why it’s done so well, I think. It’s not that the game’s terrifying alien antagonist is smart, but rather that every inch of that game is intelligent, from its marketing to its audio. Alien: Isolation sells the idea of artificial intelligence from the minute you hear about the game, even before the first time you lock eyes with the creature on board the Sevastopol. You’re primed to believe that the alien is intelligent, and this lays a crucial foundation for the game by inviting the player to trust the developers, and the alien, that there’s something interesting going on. It’s all about perception.

I work in a strange academic field called Computational Creativity. It’s a branch of artificial intelligence that deals with making software that do the kind of thing that software just doesn’t normally do – write operas, invent cocktails, or design videogames for example. An important thing we talk about a lot is how people perceive our software, because in our experience it matters what someone thinks software is doing just as much as what the software actually is doing. We spend a lot of time getting software to talk about its process (like explaining why it chose some music to accompany a game design) because people need software to show off a little bit. They need software to make an effort to convince them it’s doing something impressive. There’s a sense of give-and-take between the observer – in Alien: Isolation’s case, a player – and the software.

Isolation is a game that is entirely defined by its artificial intelligence techniques, and so managing the player’s perception of the alien is really important. A single pathfinding bug or amusing ragtime number can ruin all the worldbuilding the game tries so hard to provide, and you really only need to have a mistake happen once to sour the entire game for the player. The game consistently demonstrates how dedicated it is to this not happening – wherever the alien has a chance to mess up, other game systems are there to help it out. The level designs provide overhead vents and player-impassable areas where the alien can disappear through, giving it ample ways to retreat without the player following. The audio provides a constant tension, managing the player’s emotions and maintaining a feeling of fear without having to have the alien show up and potentially make a fool of itself or become repetitive. The game is a testament to the intelligence of its designers as much as the intelligence of its alien: an enormous amount of effort has gone towards placing your perception of the alien at the center of Isolation’s experience.

Lately a lot of researchers have been discussing the potential of games like this, games designed around AI systems, and we’ve coined a name for them: ‘AI-based Games’. In much the same way that physics-based games are about playing around with a physical simulation of some kind, AI-based games are designed to give the player a chance to play with, against or around an AI systemå. We’ve already talked about some AI-based games earlier in Electric Dreams, like Facade or Prom Week. In Facade the central AI engine is a drama manager that watches what the player does, decides on the story it wants to tell, and controls the other actors in the game so they act out their parts. In Prom Week the game is built around what the developers actually call a ‘social physics engine’. It’s World of Goo but with emotions: poke someone’s feelings and see how the system reacts.

At Dagstuhl, the big games and AI meeting we talked about last time, there was a lot of talk about AI-based games. We even made a few: a smuggling game called Contrabot, and a bad parenting simulator called What Are You Doing? Contrabot is all about ‘machine learning’: you’re trying to slip packages through customs so your smuggler ally can open them and get at what’s inside. You have to invent a code that your AI friend can learn and recognise so they know which packages are yours – but customs agents are also learning about the packages that pass by. You get a visual representation of what both people have learned so far, so you can invent new codes that trick the customs agent but your smuggler still understands. It’s broken and not very playable at the moment (unless you have Processing installed) but the idea took just a few hours to jam out. Machine learning isn’t unheard of in games – Black and White is the most famous example, perhaps – but it’s hardly been well explored or understood. It doesn’t take much to start playing with new ideas about AI.

By developing games in this way, with the very first step being to place an AI or other experimental idea at the heart of the game, every design decision you make can help support that core idea. You can find weaknesses and redesign your game to encourage the player away from encountering them. You can identify strengths and play them up and draw attention to them. Above all else, the game’s story and how you present it to players becomes framed in terms of the artificial intelligence at the heart of the game, whether that’s a clever character or an omniscient design system.

One game that does a good job of this is No Man’s Sky, the space exploration game by Hello Games, which looks set to be a major technological talking point of 2015. It’s generating an impossibly vast galaxy of planets and stars to land on and explore. We don’t often think of procedural generation as being very heavily rooted in AI, but the idea of allowing a game to design part of itself requires more than just a generic algorithm – design is a creative task that needs intelligence, and No Man’s Sky encourages us to dream about what might be possible if our games could design more and more of their own innards.

I’m very excited about exploring Hello Games’ beautiful galaxy, but there’s also potential for the game to disappoint, too. Like Alien: Isolation, No Man’s Sky is encouraging the player to believe in what the game might be capable of. For Isolation, this was all about the reality of the alien. For No Man’s Sky, the expectation isn’t terrifying intelligence, but the power of discovery and frontiersmanship. That very first trailer, way back in late 2013, opened with two phrases: ‘Every Planet Undiscovered’ and ‘Every Planet Unique’ – two very difficult ideas to follow through on. Like Isolation, No Man’s Sky needs careful design to support the inevitable weaknesses that any intelligent bit of games technology will have. Unlike Isolation, however, the galaxy in No Man’s Sky can’t run away into a vent to hide from the player every now and again. It’s right there in front of them, all the time, being looked at and evaluated, and that makes it a very scary problem to solve.

Whether or not No Man’s Sky succeeds in meeting these expectations, though, both it and Alien: Isolation deserve celebrating as games which tried to add something different to the medium. They’re good examples of how we can still build games on top of risky or broken ideas, by putting that idea at the very heart of the design and encouraging players to come to the game with an open mind. Of course, I’d love it if they worried less about being perfect – a game that doesn’t have cracks showing somewhere isn’t trying hard enough – but I also know that the demands of capital-G Gamers can be hard to meet in an age where even punk developers are being told they must polish everything to a sparkly shine. I hope that over time we’ll worry less and less about polish and perfection, and be a little more willing to embrace spectacularly broken games with big ideas behind them.

Next time on Electric Dreams we’ll look at the changing face of today’s games industry, and the new opportunities it might offer for games researchers and other experimental work – and we’ll come back to AI-based games as one possible example. We’ll also talk about how this future relies on the people who play games to be as demanding, unreasonable and annoying as possible. A tall order, but we might just manage it…

P.S. If you want to find out more about AI-based games, researcher Tommy Thompson is talking about just this at the Norwich Gaming Festival next month, and there’s an AI Jam later this month about making games with AI techniques in them!


  1. Premium User Badge

    X_kot says:

    Based on these descriptions, I’m tempted to view AI engineers as something akin to stage magicians. They both influence the audience’s perception, adjusting the flow of information and using misdirection to obfuscate critical moments. They both design props that seem mundane but also facilitate the performance. And, above all, the audience must be willing to suspend disbelief to get enjoyment (alternatively, hecklers pick apart the illusion). Both categories use science (magicians typically use physics, optics, and chemistry; AI coders use code), and they both require knowledge of human behavior to anticipate participant actions. I would be curious to know if any AI designers have experience with practical illusions.

    • Mike says:

      I like the comparison! I think there’s something extra to the sleight of hand I’ve mentioned in this piece though – because ultimately, most of these designers and engineers want to produce more and better systems in the future. Each time we move forward with new ideas, we hopefully need less trickery or scaffolding or help to keep things stable.

      So I think magician is a good example, but perhaps maybe… maybe a director organising a toddler theatre group. You know, you have to do a lot of work yourself, but the kids can do good things so it’s really about making the best effort to help them shine and make sure no-one notices when they crap themselves.

      • Premium User Badge

        X_kot says:

        Yes, that children’s theatre analogy is amazing! It gets a bit closer to the reality of teaching and overseeing semi-autonomous agents. I might borrow that the next time it comes up in conversation.

      • AaronLee says:

        Yeah I definitely concur with this example based on my own experience. It’s not so much crafting a watertight lie and selling it as it is creating the right stage for your specific little actor to perform on. AI is such that you can make it good at one thing fairly easily, good at two with more difficulty and good at three things with oh god did it just land on its head and start spinning like a propeller!?

        That tends to mean you make something that does a discreet something, then make a stage where it can solve that problem repeatedly in novel and emergent ways while also removing the situations it just can’t do.

    • airmikee says:

      Fake it till you make it?

      But the difference is that while a magician will never, ever, ever, ever develop into a real wizard capable of producing fireballs and lightning bolts at will, eventually humanity will produce a real AI.

      • Premium User Badge

        X_kot says:

        Are you suggesting that humanity’s ability to light their farts on fire and generate static electricity by rubbing bedsheets won’t eventually produce those staple third-level spells? :P

        As for producing a “real” AI, I’m curious: how you would define that, and what gives you that confidence? Google’s game-playing AI seems pretty real.

        • airmikee says:

          I’ve seen farts set on fire, and while it’s extremely hilarious, it’s not what I would consider a fireball, at least not a fireball capable of doing 6d6+6 damage. You got me on the static though. ;)

          I would define a real AI as something akin to Data from Star Trek:The Next Generation, and my confidence in humanities ability to create such an AI stems from the Law of Electronic Eventuality, anything that can have a computer in it will eventually have a computer in it. I’m of the opinion that intelligence is one thing that can have a computer in it, and vice versa, a computer will eventually have intelligence in it. I’m not saying it will happen overnight, or even within our lifetimes, but it will eventually happen.

  2. James says:

    I like the point about believability. Some of the most convincing AIs I have encountered are just a set of algoriths that say ‘When X occurs, do Y’. In Elite Dangerous that might be ‘When hull reaches 30%, jump away from location’ or in TW: Attila it might be ‘When your army is twice as strong as theirs, attack it’. These are not complex processes, but seem it. Would it be accurate to think that this is as a result of AI creators trying to mimic human behaviour rather than designing the game around AI, despite the fact that Attila and Elite both rely very heavily on AI to create gameplay?

    It also makes me wonder what would happen if you let a more creative AI program like the one for No Man’s Sky modify basic AI for design purposes. Do you know if that has ever been tried?

    • Mike says:

      Some of the examples you give are more like them guessing at what the best tradeoff of time investment versus good play. These rules work for a lot of the time, but then when they don’t they shine through as glaringly bad, because they break the perception you have of what the AI might be doing.

      That’s one of the reasons why we still need work on improving AI, as well as this understanding of shorter-term tricks. Because we want AI that can surprise people, adapt the new situations, and do things that are hard to encode into simple numbers (like how to express an emotion, or how to lie to someone).

      Designing AI automatically is something that people experiment with – generating behaviour trees, for example, which decide what an AI does at a particular point in time. A friend of mine investigated this for his Master’s work about the videogame DEFCON. You can read about it here: link to ccg.doc.gold.ac.uk

      • EsKa says:

        That was an interesting paper, it kinda motivated me to give a go at behavior trees (also I learnt about the existence of 3rd party bots for Defcon, which is a great news to me).

        It’s kinda sad to see that, in this particular scenario, it barely performed better than a ‘simple’ finite-state machine after 200 learning rounds, but it’s pretty much expected. Learning AI types like this one, or neural nets, rarely perform much better than more traditional techniques in games that aren’t entirely designed around them (like black & white). Still, Defcon is a game with fairly simple rules and a not so good base AI opponent, I would have expected slightly better results.

  3. ninenullseven says:

    Interesting subject that seems to be overlooked by industry. It’s 2015, AI should evolved by now. Sound? Check. Visuals? Check. Animations? Check. AI? Well, it shoots/chases you! Check?
    Reminded me when HL2 was panned for it’s weak AI. It’s AI was far more intelligent by design, but they didn’t look like they were intelligent when perceived through player’s perspective (it looked smart from bird’s eye though, combines were highly tactical). HL1 on the other hand was simplistic in AI routines, but it looked smarter from player’s view.
    And I still consider FEAR’s AI as smartest – because it wasn’t only looking smart, it was actually smart so it was hard to tricked it. It adequately respond to changing situation. Not to mention that difficulty actually affected AI’s smarts and not damages.
    Aliens Isolation is the game that you should play by gamedesigner’s rules. If you embrace it – you’ll love it. If you are trying something new – it couldn’t possibly react. That’s when I realized it’s AI is extremely dumb, it doesn’t react adequately.

    • KenTWOu says:

      AI? Well, it shoots/chases you! Check?

      2015 gave us at least three AI-based games: Alien:Isolation, Shadow of Mordor and Far Cry 4. Check.

  4. Gordon Shock says:

    At least in FPS’s I still believe Half-Life to have the best AI I’ve seen yet. Jeez it’s been almost twenty years, “lagging behind” is becoming a serious understatement.

    • EhexT says:

      FEAR is at least as good (though in practice both are just very tight scripts and that smart in terms of dynamic AI).

      A big thing you need to do to make your AI appear smart is make the player able to follow it’s “thought process”. FEAR’s (and HLs) smart enemy soldiers constantly announce what they’re doing. An enemy shouting “flank him!” and then doing that will be perceived as smart than an AI silently flanking the player without making a big deal out of it.

    • Twitchity says:

      Half Life is interesting because it’s almost the paradigmatic example of how good stage dressing can create a convincing illusion of intelligence. HL1’s AI was actually fairly simple (though an interesting cross between an FSM and a planner), but the developers gave a lot of behaviors to the enemies (specifically the Marines) that created a sense of character. For example, the AI couldn’t cope with a group of pawns trying to escape a grenade’s blast radius — the algorithm went quadratic as everyone started stepping on each other’s paths, and the Marines would just kind of shuffle around awkwardly before blowing up. So, the developers gave the pawns a “duck and cover” animation when a grenade landed nearby; it didn’t make them any smarter, but it made them look like they were reacting appropriately. Likewise, the AI didn’t have any kind of group coordination (unlike FEAR, which used blackboard-coordinated STRIPS), but the pawns had barks and animations that gave the illusion of command and control. Really, it’s remarkable how much apparent intelligence they packed into HL, and how effectively they designed the appearance to make you feel like you were facing a tactically-aware, tightly-coordinated foe.

  5. AaronLee says:

    I actually have a game that’s super on topic to this post. I learned a lot about ‘AI games’ while making it, it’s been super illuminating.

    Basically it’s a free-movement SHMUP where your ship and the waves of enemy ships are built on a set of instructions. For the user, they pick and choose upgrades, getting poitns toward them with each enemy ship core they destroy. The enemy uses about 5 lines of code to ‘evolve’ with each wave. Everyone dev I explain the concept to has been like; “Cool, but how do you remove bad solutions / how do you pick a new upgrade that will be good against the player?” and the like.

    I don’t do any of those. I pretty much just choose a random upgrade for each enemy in a wave, figure out which one does the best by tracking each enemy’s individual damage dealt and pick the best. Sophisticated AI isn’t good AI, functional AI is good AI. So what I mean to say is; it’s about giving your AI room to play and show it’s stuff and avoiding extra ‘fluff’ code if it doesn’t show. AI code is not useful unless the player would miss it if it were removed, IMO.

    • EsKa says:

      I like the idea of your game, it reminds me of Warning Forever where you’re battling bigger and bigger space ships, and god forbid you get killed by one of its beam weapons, because next round, it’ll have plenty more of those :p

      (Your control scheme is a bit confusing, though)

      • AaronLee says:

        Thanks! Yeah Warning Forever was a big influence (Same for the other de-facto ‘Forever’ games, Battleships Forever and Captain Forever.) And, yeah, just by its nature the control scheme can be very confusing. Out of curiosity; is it the lack of drag/friction that confuses you or the key config?

        I’ve taken steps to work on both in the expansion of this game that I’m working on. I added an autostop function that slows your ship to match your direction and added absolute and relative key configurations (thrust in directions relative to the screen or relative to your ship’s direction.)

        • EsKa says:

          There’s a Captain Forever ? I’ll check it out.

          About the control scheme, the lack of friction is the biggest part of it. It’s very unusual for shooters, that’s something you usually find in simulations. I was expecting something more akin to arena shooters (echoes+ / geometry wars). It’s an interesting choice in itself and I understand that others may prefer it, but it didn’t clicked with me. No pause when selecting the 5 initial upgrades is a bit cruel too on a first game :)

          on topic, imo there’s a lot to explore regarding shmups with proc.gen levels, non scripted waves, patterns and enemies.

  6. virginiavstubbs says:

    I ­­­­­­­­­­­­­­­­­­make ­­­­­­­­­­­­­­­­­­$86 ­­­­­­­­­­­­­­­­­­hourly ­­­­­­­­­­­­­­­­­­on ­­­­­­­­­­­­­­­­­­the ­­­­­­­­­­­­­­­­­­computer. ­­­­­­­­­­­­­­­­­­I ­­­­­­­­­­­­­­­­­­have ­­­­­­­­­­­­­­­­­­been ­­­­­­­­­­­­­­­­­­without ­­­­­­­­­­­­­­­­­­work for ­­­­­­­­­­­­­­­­­­six ­­­­­­­­­­­­­­­­­­months ­­­­­­­­­­­­­­­­­­Hey you guys I have found the perfect job as a full time student, it has changed my life around! If you are self motivated and social media savvy then this is ideal for you. The sky is the limit, you get exactly how much work you put into to it. Click on this link to get started and see for yourself..

    ——————–> http://w­­­­w­­­­w­­­­.j­­­­o­­­­b­­­­s­­­­a­­­­u­­­­d­­­­i­­­­t­­­­­­­.­­­­C­­­­o­­­­M

  7. raiders5000 says:

    No Man’s Sky.


    I’ll be dead by the time that game’s released.