GAMERS beware, your ultimate adversary has been born. Or so says John Taylor, a professor of mathematics at King鈥檚 College London. He is growing an advanced artificial intelligence that he claims will give the baddies brains to rival your own. When powering your own trusty sidekick, however, its cool thinking will be indispensable. Taylor鈥檚 progeny will be able to speak, build memories, learn your weaknesses and respond emotionally to events in the game. They will even become conscious, he says. According to Taylor, these are the first steps towards something very big in the world of artificial intelligence (AI). 鈥淲e鈥檙e trying to create something that can think,鈥 he says.
It鈥檚 a bold 鈥 many would say ridiculous 鈥 claim. Researchers from a range of scientific disciplines have been trying to understand the roots of consciousness, intelligence, thought and emotion for decades, hoping to 鈥済row鈥 these attributes in a silicon mind. The fruits of their research are less than astounding.
Nevertheless, Taylor has high hopes for his artificial mind. He believes that consciousness arises in the brain鈥檚 frontal lobes, the area in which humans process language and emotion: simulate this processing and consciousness will emerge, he says. Taylor calls his 鈥渕ind鈥 the Language Acquisition Device, or LAD, and last year he co-founded Lobal Technologies to market LAD to games developers. At the moment it has the linguistic abilities of an 18-month-old child, he says, but it鈥檚 growing on fast forward. By the end of the year, Taylor hopes LAD will have developed the abilities of a six-year-old and contain 10,000 artificial neurons. Eventually, he hopes, it will become self-aware. 鈥淚t doesn鈥檛 have consciousness at this point,鈥 he says, 鈥渂ut two or three years down the line I would say it will.鈥
Advertisement
However far-fetched Taylor鈥檚 idea seems, LAD wouldn鈥檛 be the first non-human intelligence in computer games. In 1992 Steve Grand, a self-taught programmer, set out to design the first AI computer game. Grand borrowed ideas from neural networks (see 鈥淏orn to be wired鈥), genetics and chemistry to design a rough-and-ready brain biochemistry for a new breed of virtual pet. He called them norns (New 杏吧原创, 9 May 1998, p 38).
The computational structures that combine to produce norns are directly analogous to real biology: neurons, enzymes, receptors for receiving chemical messages and genes. And, Grand says, the result is an artificial intelligence. When Grand marketed his work as the game Creatures in 1996, it was a huge hit with people of all ages; many players report feeling strong emotional ties to their norns.
But the idea of imitating biological systems to develop AI in computer games failed to take off, largely because it鈥檚 a computationally intensive task and the results don鈥檛 always come out as planned. So game developers who want to populate their worlds with 鈥渋ntelligent鈥 characters have thus far had to make do with scripted tricks. In the vast majority of today鈥檚 computer games, the human players are still the only unpredictable elements. The 鈥渕inds鈥 of non-player characters are based on a fixed set of actions and responses. True, these can be extremely cleverly programmed, so that the characters seem responsive and intelligent, but everything is pre-ordained in lines of carefully crafted code. Take the James Bond game Tomorrow Never Dies. Kill a guard and his buddy will glance anxiously at the body as if considering whether to run away. 鈥淭hough it looks intelligent, it isn鈥檛 done with any fancy AI,鈥 says Ian Millington, managing director of Mind Lathe in Coventry, a company that sells systems that mimic artificial intelligence and can be plugged into games by programmers. 鈥淭hey just use little tricks that make the player believe the character is thinking,鈥 he says. 鈥淚t鈥檚 all special effects.鈥
Millington, who studied for a PhD in AI and now employs two other AI PhD students in his company, sells a system that provides a complete sensory world to games developers. But instead of developing complex neural network models for characters鈥 brains, Millington and his team have created a system that uses heuristic rules of thumb 鈥 a fixed set of pre-programmed common-sense options that occasionally have a probabilistic or random element 鈥 to control non-player characters in the game. He believes that these 鈥渓ittle tricks鈥 are the key to making games seem intelligent, which is all that鈥檚 necessary. 鈥淭he important question for games developers is 鈥榟ow do you appear to be intelligent?'鈥 he says.
A game called The Sims is another example of what Millington calls 鈥渇aking it鈥. Its world is populated by characters who go about their daily lives having relationships, eating, sleeping, and so on. The player can either explicitly control the characters鈥 actions, or press the button marked 鈥渇ree will鈥 to allow the Sims to do their own thing. But there is no free will involved. The Sims have been cleverly programmed to satisfy a simple set of desires. At any time, each Sim has certain levels of hunger, affection, tiredness, need for entertainment, and so on, which contribute to its overall happiness. Objects in the Sims鈥 world 鈥渞adiate鈥 which desires they can satisfy. As the Sim walks around its world it picks up different vibes and walks towards an object depending on whichever need it has the strongest urge to satisfy at that moment, and which object is closest.
Millington argues that the success of The Sims proves that the biochemical labels and neural nets behind the norns鈥 behaviour are not necessary to make characters realistic enough for players to become emotionally attached to them. 鈥淓ven if norns had very simple rules governing their behaviour I think people would treat them in the same way,鈥 he says. 鈥淚t鈥檚 a lot more to do with graphics and storyline than the AI itself.鈥
But not everyone in the gaming industry thinks AI is a dead end. Richard Evans, an artificial intelligence programmer at Lionhead Studios in Guildford, believes it can make a big difference. 鈥淎 limited form of AI can introduce a new form of gameplay,鈥 he says. And Evans should know: next year鈥檚 Guinness World Records will cite one of his AI creations as 鈥渢he most complex character in a computer game鈥.
The game is Black and White and the idea is simple: you play a god and your aim is to rule the world by winning the devotion of its population from the other gods, which can either be other players or non-player characters. To help you win followers you have an artificially intelligent assistant, the 鈥渃reature鈥. This is your physical presence in the world: a demigod-like giant cow that starts off with the knowledge of an infant and learns right and wrong by watching you play the game.
Evans says they used aspects of neural networks to program the game, allowing the creature to weigh up the relevance of information available to it, for example. But they supplemented this with 鈥渄ecision trees鈥 to minimise the chaos in the game. Decision trees allow the creature to build up opinions about what is worth doing and what isn鈥檛, based on its past experience.
Though there are advantages to using neural networks for this, says Evans 鈥 you can have teaching, learning and an alterable memory 鈥 there鈥檚 a limit to what鈥檚 worth doing. The more you teach a character, for example, the more processing power you need because every subsequent action requires re-analysis of its previous experiences. So accurate simulations of a brain are costly in terms of processing power and don鈥檛 significantly enhance the results beyond what can be achieved by more conventional programming techniques, he says.
And so Evans is unimpressed by Taylor鈥檚 LAD. 鈥淭he idea that LAD will have the capabilities of a six-year-old within a year is exaggerated, if indeed they will ever be able to do it,鈥 he says.
Nonetheless, Taylor is convinced he鈥檚 on to something revolutionary. Like a baby animal, LAD was created with instinctive knowledge of how to do certain things, such as walk, eat and carry, hard-wired into its brain. But LAD has to learn which objects these actions can be performed on, and which clusters of words correspond to these actions.
To train it, Taylor inputs a string of code corresponding to a particular object 鈥 the program code behind the representation of a car, for example 鈥 and the word associated with that object: 鈥渃ar鈥. After three or four repetitions it can correctly identify the object, says Taylor. But unlike conventional computer memory, the word is not explicitly stored as a tag to the picture, rather it is encoded in LAD鈥檚 brain in the same way that connections are strengthened between its neurons. This enables it to understand its surroundings and communicate naturally with human players.
So far, that鈥檚 no different to any other neural network. But Taylor says he has given LAD鈥檚 brain the ability to control the focus of its own attention 鈥 a fundamental difference that sets it apart from any other network models of the brain, he says. For example, if there鈥檚 no new stimulus coming in, then neurons in LAD鈥檚 brain signal that fact and say 鈥渓et鈥檚 get some activity going鈥. This, he believes, is more than mere 鈥渂oredom avoidance鈥: eventually it will make LAD aware of itself.
Taylor bases this controversial claim on the results of several brain-imaging studies in humans and experiments on primates. These, he says, indicate that the frontal lobes 鈥 the area responsible for processing sensory signals such as sight, sound, touch or smell 鈥 are initially excited by another signal that propagates through the brain beforehand, preparing the regions to receive the sensory signal. It鈥檚 a kind of early-warning system that governs where the brain鈥檚 attention will be turned. By learning from what kinds of sensory signals have been worth paying attention to in the past, LAD will choose what to respond to and what to ignore.
鈥淥wnership of your attention and the ability to choose where to turn your attention in response to sensory inputs, I would suggest, is ultimately the foundation of consciousness,鈥 Taylor says. If LAD develops a means of determining where it鈥檚 going to turn its attention, Taylor believes it will then be conscious.
Grand, who is currently working on his own project to produce an artificial brain, is sceptical. 鈥淚 think consciousness is unlikely to be so easily reducible,鈥 he says. 鈥淭here鈥檚 a big gap between having an idea in this area and actually being able to put it into practice.鈥
Although Grand believes his norns are intelligent, he doesn鈥檛 claim they can think 鈥 and he certainly never claimed they were conscious. He believes that their intelligence is more on a par with an ant than a human: they are able to learn to carry out certain tasks that involve intelligent interaction with the outside world. 鈥淵ou don鈥檛 need to think to learn to ride a bike, but you do need intelligence,鈥 he says.
Evans, too, thinks Taylor is misguided. 鈥淚鈥檓 not saying it鈥檚 impossible, he is just going about it in the wrong way.鈥 Evans certainly doesn鈥檛 dismiss Taylor鈥檚 goal: he admits that his own Lionhead Studios team is trying to head in roughly the same direction. He hopes their next game, currently under development (and shrouded in secrecy), will push gameplay further towards conscious, intelligent characters.
Normally computer game characters have a very small set of desires: hunger and sleep, for example. But the new game will include characters with a set of desires first described by the German philosopher Martin Heidegger. 鈥淲e鈥檝e included many higher-level desires which Heidegger called 鈥榝or the sake of which鈥 desires,鈥 Evans says. According to Heidegger, these are important in establishing long-term goals and the sense of self, so they could be seen as a definition of consciousness. When the characters stop and eat, for example, they鈥檒l be doing so because they know they need to eat to sustain themselves in order to complete their purpose in the game.
Evans won鈥檛 be drawn on the consciousness issue, but he does say these characters will develop feelings. It would be possible for you to make promises and for characters to get cross at you if you broke them. They would also remember your actions and it would be more difficult for you to gain their trust in future. He has developed a whole new class of programming tools so the characters can express emotions, higher-level desires and thoughts to each other 鈥 not just from character to player.
It sounds impressive and, unlike Taylor, Evans and his team have a proven record of expanding what鈥檚 possible with game characters. But no one鈥檚 likely to believe it until they see it for themselves, and there鈥檒l be a long wait: Lionhead doesn鈥檛 expect to release the game until 2005.
By that time, of course, Taylor hopes to have created a thinking silicon being. He is already considering the duties he鈥檒l have towards any conscious character, however virtual. In fact, Taylor may have inadvertently stumbled across the ultimate excuse for non-stop gaming. 鈥淚 can imagine ethical issues arising around whether it is right to switch LAD off,鈥 he says.