杏吧原创

Also see: What counts as a conscious thinking machine?

THE warrior dashes through one dark corridor after another, as laser shots crackle past his body. He crouches, jumps, swivels鈥� does he detect the source? He fires at a potential enemy, barely visible in the shadows, then decides to flee.

Though the character inhabits the first-person shoot 鈥檈m up video game, Unreal Tournament, there鈥檚 no human player in the driving seat. This warrior, called , is a character controlled entirely by a biologically inspired model of consciousness. The feat could help us to build more human-like machines, and even shed light on the workings of consciousness itself, one of the biggest mysteries in science.

The thrilling run I watched was just a practice, but next week the neuroscience-inspired warrior will compete in the fight of his life: the annual . It is a video-game alternative to the Turing test, the most famous test of machine intelligence.

In the Turing test, programs called bots try to convince human judges that they are human via text-based conversation. In the BotPrize contest, multiple humans and bots play Unreal Tournament simultaneously, all-against-all and anonymous in the same arena of battle. At the end of the match, the human players judge the 鈥渉umanness鈥� of all of their opponents.

The winning bot will be announced at the in Granada, Spain, on 13 September. The prize is $7000 shared between all the bots that receive as many 鈥渉uman鈥� votes as the average human player. If none reaches that threshold, the bot with the highest humanness score wins $2000. So far, bots have only won the smaller prize. An earlier version of Neurobot finished second last year, after technical difficulties hampered it in the final round, says its creator, at Imperial College London. He expects to go one better this year.

It鈥檚 difficult to unpick exactly what gives a bot the human touch in Unreal Tournament, but Philip Hingston of Edith Cowan University in Perth, Western Australia, who has run the contest since 2008, has some ideas. 鈥淪ome of the judges have reported the feeling that one of the players is determined to get them 鈥� some kind of sense of purpose,鈥� he says. By contrast, bots seem less likely to focus their attention.

The idea behind BotPrize is to foster software capable of navigating physical space in a human-like way. This could be used to create more realistic video-game characters, better simulate crowd behaviour in emergency situations or control robots in the real world.

Neurobot鈥檚 performance, in particular, will provide an indication of whether the theory of consciousness that it is based on 鈥� global workspace theory (GWT) 鈥� can really produce human-like behaviour.

According to GWT, unconscious processing such as the gathering and processing of sights and sounds is carried out by various autonomous brain regions working in parallel. Only when information is deemed important enough can it enter the global workspace 鈥� or 鈥渃onsciousness鈥� 鈥� and be broadcast to other regions of the brain.

Though a program that mimicked a simple version of GWT in software 鈥� without virtual neurons 鈥� won BotPrize in 2010, Neurobot is the first to implement it from the bottom up. It simulates 20,000 individual neurons and the electrical currents that flow between them, creating 1.5 million connections. This is a far cry from simulating a whole brain 鈥� made of about 120 billion neurons 鈥� but it is an important step if such bots are to inform theories of consciousness, says Murray Shanahan, Fountas鈥檚 supervisor at Imperial.

To implement GWT, Neurobot鈥檚 simulated neurons are divided into different populations, each responsible for controlling and reacting to different actions and perceptions. One set is responsible for sensory information, for example, so produces spikes in electrical activity in response to objects moving into Neurobot鈥檚 field of view, or incoming shots from enemies. Another set is responsible for Neurobot鈥檚 behaviours, and includes subsets such as movement, navigation, exploration and gun-firing.

These populations work simultaneously as Neurobot goes about its business, but at any one moment, only one can broadcast from the bot鈥檚 global workspace. Like actors on a stage vying for the spotlight, which process ends up in the global workspace depends on which neurons have the highest prominence at any given moment. Winning access means those neurons can communicate with others. For example, if the bot is to approach an object, information must be shared between the set of neurons that perceive the object and behavioural neurons that move its body. This can only happen if one of these sets makes it into the workspace.

At the same time as this focused behaviour is going on, there are other actions that don鈥檛 require Neurobot鈥檚 attention. These can be likened to the way that people in conversation walking down the street avoid obstacles effortlessly.

It is this mixture of conscious and unconscious processes that Fountas hopes will make Neurobot鈥檚 behaviour seem human. 鈥淚 have no idea why he chooses to shoot when he does,鈥� Fountas says. 鈥淭he decision is completely his.鈥�

鈥淚 have no idea why NeuroBot chooses to shoot when he does. The decision is completely his鈥�

Neurobot鈥檚 main rival, ICE-CIG2012, created in the Intelligent Computer Entertainment Lab at Ritsumeikan University in Kyoto, Japan, approaches it differently. That bot comes preprogrammed with tactics it has learned from previous human players. Last year, that strategy allowed ICE-CIG to beat Neurobot to the prize. Five other teams will also be this year.

Whether it wins or not, the most intriguing question of all remains unanswered: is Neurobot actually conscious?

鈥淚t鈥檚 a dangerous word,鈥� says Fountas. 鈥淚 hope he is.鈥�