This is very much a scene-setter, setting out my views about how virtual worlds and AI interact, and the key role that virtual worlds might play in the development of AI. Think of it as the introduction to the thesis.
This section, in modified form, has been accepted for the AI-2008 Twenty-eighth SGAI International Conference on Artificial Intelligence CAMBRIDGE, ENGLAND 9-11 DECEMBER 2008
Abstract The last two years have seen the start of the commercial activity within virtual worlds. Unlike computer games where Non-Player-Character avatars are common, in most virtual worlds they are the exception – and until recently in Second Life they were non-existent. However there is real commercial scope for AIs? in these worlds – in roles from virtual sales staff and tutors to personal assistants and “ghosts”. Deploying an embodied AI into a virtual world offers a unique opportunity to evaluate embodied AIs?, and to develop them within an environment were human and computer are on almost equal terms. This paper presents the architecture being used for the current deployment of chatbot driven avatars within the Second Life virtual world, looks at the challenges of deploying an AI within such a virtual world, the possible implications for the Turing Test, and identifies research directions for the future.
''...you know he claims that he is not an AI. He says he passes the Turing Test every day, but he is not intelligent: just a sequence of yes/no responses, just a massive algorithm Divergence, Tony Ballantyne, 20078''
In the half-century since the Turing Imitation Game29 was created technology has changed dramatically. The paper, card and text interface was replaced by the interactive terminal, and then by the desktop PC with the windows, icon, mouse and pointer (WIMP) interface. Technologies such as Flash and text-to-speech have enabled us to create avatars on the desktop or on web pages – human like characters which can attempt to converse with users when controlled by a chatbot34 programme. The most well known chatbot is probably Anna on the Ikea site16. However chatbots have never really caught on, possibly partly due the immaturity of the chatbot engines, but also due to the way that the conversation model breaks the WIMP metaphor of the rest of the interface.
The last couple of years though have seen the emergence of a new interaction model – the virtual world. Here the computer creates a complete 3D environment, and the user, represented by their own avatar, can move around the 3D space, meet and interact with avatars controlled by other users, and change and build new environments and new devices. Linden Lab's Second Life30 is probably the best current example of an open virtual world – one in which users have almost as much freedom of action as they do in the real world. Second Life grew from 100,000 registrations in Apr 2006 to over 13m registrations by Apr 2008. More significantly organisations ranging from the BBC and IBM to Save The Children Fund and the British Council have started using virtual worlds both within and external to their organisations.
Virtual worlds themselves partly grew out of Multi-User-Dungeons (MUDs?), Massively Multiplayer On-Line Role-Playing Games (MMORPGS) and computer (and even paper) role-playing games. In all of these the “Non-Player Character” (NPC) has always been present. The NPC is a character (or avatar) controlled by the game (computer) and which is used either to impart knowledge or things, act as friend or foe, or just provide local colour. Whilst their scope to have an open ended conversation has been limited (and usually not even present), the fact is they were designed to blend in with the human avatars and the computer generated environment.
In virtual worlds such as Second Life the NPC has been more or less completely absent. Partly this was due to an initial resistance to such characters from Linden Lab (this was after all meant to be a shared virtual world where human to human interaction and socialisation were paramount), and partly since technical limitations (lack of avatar control, lack of web interface or powerful programming language) made it hard to create even a basic NPC.
Now, however, these technical limitations have gone, and Linden Lab is taking a more open view. It is also worth noting that some competitor virtual world platforms (such as Active Worlds2 and Kaneva17) offer NPCs? “out the box”. The result is that we are now able to fully experiment with chatbot controlled NPCs? within a virtual world. The roles that such an NPC can play within a virtual world range from virtual receptionists, greeters and guide, to personal assistants, mentors and tutors, and ultimately perhaps as personal agents (or “ghosts”) – controlling a user's avatar in their absence – or even death.
Chatbot development is reasonably well studied ever since the TIG was first proposed. ELIZA13? was the first famous chatbot, and ALICE6? was another milestone. The Loebner Prize20 and the ChatterboxChallenge9? are both annual competitions which have their roots in the TIG.
However these are typically text-only experiments – although some limited visual components are often added – the focus is on whether through the text exchange alone we can replicate human “behaviour”. However with virtual worlds we have the ability to embody the chatbot. The new challenge is:
“Are we able to create an NPC within a virtual world which is indistinguishable in its complete behaviour from a player character/avatar”.
And if we can do so, will we have passed the Turing Test?
We try to keep away from using the term Artificial Intelligence since there appears to be no commonly agreed definition of what Artificial Intelligence1 is, the term “AI” brings with it the grand visions of science fiction of powerful artificial intelligences, and more founded concepts such as the Singularity19, and in academic parlance the term AI is now being replaced by Artificial General Intelligence15 (no doubtly partly for the two reasons above). The defining characteristics of AGIs? appear to be around problem solving25, learning25 and coping within insufficient information33.
We are under no illusion as to what we are trying to make or study – we simply aim to computer programmes which mimic human behaviour. As such we prefer to refer to our creations as either Robotars (an avatar in a virtual world which is controlled by a computer, rather than a person), or Personality Constructs (PCs? - Robotars with a consistent human-like personality). Neither make any special claim to their “intelligence”, they could be very simple, or ultimately very advanced.
The Turing Test (or a virtual world version) is however still a good milestone for PC development. Indeed, for perhaps the first time within the context of the Turing Test, virtual worlds place the human and the computer on an equal footing10. Both are operating at a level of abstraction beyond their “normal” environment. Both are operating as avatars in a virtual world (and until recently both were constrained to text chat). As such the computer is finally presented with a level playing field in which to take the Turing Test (although such a window may be closing as voice gets introduced to virtual worlds).
A significant aspect of a virtual world such as Second Life (but interestingly not in a MMOPRG like World of Warcraft36) is that the working assumption that a human player has when the encounter another avatar is that the other avatar is also being controlled by a human. In a conventional Turing Test the challenge that a computer has is to prove that it is a human. Within a virtual world the challenge is subtly different – it just must not give-away the fact that it isn't a human.
Whilst developing PCs? as a purely academic exercise has its attractions, we are a commercial organisation. This means that our developments are focussed on creating chatbots with commercial uses. For instance we are already involved in deploying chatbots as:
To us the most immediate test of a chatbot's salience is the satisfaction of the customers using it. This may or may not have any correlation with a bot's Turing performance.
Since late 2007 we have deployed two robotars within the Second Life virtual world. Abi is our virtual receptionist. She staffs our office 24/7, greeting visitors and answering basic questions about Daden and what we do. She also knows her way around the office area. In fact Abi has two simultaneous existences, one as an embodied avatar in Second Life, and one as a head-and-shoulder Flash based avatar on our web site. Both share the same chatbot engine. Halo is our PC test-bed. She has been running on the web for over 4 years now, and now has a Second Life presence. We are building her her own “home” in Second Life, starting with a garden, and she is being given a far more open set of goals in terms of what she can do – ultimately to be driven by her own motivation model. In the last 3 months Abi has had 1260 conversations, of which about 140 were in Second Life (the rest being on the web), and Halo has had 840 conversations, of which 32 were in Second Life.
Within our Chatbot system we have taken a very pragmatic approach drawing on technologies such as:
Our chatbot engine into which we are incorporating these features is called Discourse11.
Although we have experimented22 with mimic based automated learning systems (such as used by the George chatbot14), we are not finding this suitable for the type of chatbot/NPC that we are trying to develop (i.e. one which is trying to fulfil a purpose other than just engage in an open ended conversation with no real goal or meaning).
In developing a technical architecture for embodying our chatbots within a virtual world we were guided by a number of principles:
That the robotars appearance and interaction in the virtual world should technically be the same as for a human controlled avatar
Figure 1: Embodied Chatbot Architecture for Second Life
The architecture comprises 4 main elements:
It may be that existing schemas such as RRL27?, the related Gesticon18, or any of the Virtual Human Markup Language26 (VHML), although the former seem to complex for our needs, and the latter appears to have stagnated. To deploy the same bot into a different virtual world should then only require us to change the interface code (the element represented by libsecondlife and chatbotif in this example).
From our first deployment of these avatars into The virtual world offers a number of new challenges to the chatbot and AI designer. Whilst some have been encountered in the past in work on robots, others are new. Particular issues which we have already encountered and are now seeking to investigate, include:
6.1 Handling Group Situations Most prior work on chatbots has been on the one-to-one situation, one person talking to one chatbot. Very early in our virtual world work situations arose where two human avatars would start talking to one robotic avatar/chatbot. The chatbot, knowing no better, would try and respond to everything it heard, bot just those things directed at it. How we make chatbots cope with a group situation is a major challenge.
6.2 Multi-line input Traditional chatbot testing involves the chatbot responding to the input as the ENTER key is pressed. However in a virtual world a user will typically enter a long statement as a series of individual lines of text. The bot needs to identify when the person has stopped entering text and the bot can start working out its reply (although even humans find this a hard task when they first enter a virtual world)
6.3 Movement NPCs? in game environments typically “know” the layout of the terrain from privileged access to the world “map”. In virtual worlds the bot both has no such privileged access (it only get the same location data as is visually presented to the human user), and the landscape itself can be changed at will be suitably authorised users (and even bots themselves). We have initially dealt with this by manually creating a topological map of way-points with unobstructed routes between them, and then applying a least-distance algorithm to find a route between any two waypoints. However at the very least the bot should also run some form of collision avoidance/recovery software to cope with changes since the map was made, and ideally should navigate purely by a set of physical robot style route-finding algorithms.
6.4 Expressing and Identifying Emotion and Gesture Web based chatbots have typically had limited ways in which to express emotion. Embodied chatbots are able to express emotion through facial expression, limb and full body gestures, and even movement (eg changing social distance or running away). We need to find a way to code, cue and control all of these. In addition the bot should be able to recognise the body language cues being given off by the avatars around it – reading social distance and gesture. A challenge here is to choose an approach which works at an “engineering”, gross gesture, level, and which can ideally be deployed in multiple worlds through something like the AAML/ASML languages mentioned above. We are currently looking at models such as Ekman12 and OCC24?, and recent work at the University of Wolverhampton31.
6.5 Object Identification, Location and Interaction Robotars have the same ability to interact with objects in the virtual world as do human controlled avatars. At the very least a bot needs to know how to sit on a chair (and find one to sit on), recognise significant features in its environment (eg a swing or a desk), and be able to exchange objects with other avatars (eg handing out flyers or gifts, and even receiving payments). Again this should be achievable through a suitable ASML/AAML implementation.
6.6 Memory A bot is exposed to far more data, and more diverse data, than any chatbot. Any avatar moving near it will generate a stream of location data – even more if gesture is being captured. Any change in the environment generates more data. Every step the bot takes yet more. Whereas storing every conversation that a chatbot has is relatively trivial trying to store and index the environmental data, particularly in a format like RDF which will let the bot analyse the data later is a significant challenge.
6.7 Motivation Although our bots find themselves having several conversations in the virtual world each day the fact is that they spend most of their time doing nothing. When a bot is just an instance which is created when someone wants to talk to it, and destroyed once the conversation is over this does not seem an issue. But when the bot is visually there, standing and metaphorically (or literally) twiddling its thumbs it is natural to think about how the bot should fill its time. Whilst such time-filling activities could be explicitly scripted it seems more in keeping with an embodied avatar and personality construct to give the bot a suite of possible actions, a set of needs and some form of motivation model to let it decide what to do and when. Our current thoughts are that such a model could be based on something like Maslow's Hierarchy of Needs21, and that actions could range from learning its environment and reading RSS feeds, to finding people to talk to and learning the best ways of earning money for itself.
To us the most interesting aspect of this whole field is how the Turing Test relates to Personality Constructs in virtual worlds. Quite apart from the more theoretical aspects already mentioned (e.g. the level playing field) there are several practical ways in which a Turing Test carried out within a virtual world may be more “valid” than one conducted through a character based interface. There are at least three areas worthy of consideration.
First, the additional cues which the NPC can generate to try and prove (or not disprove) its human-ness such as:
This is in addition to some of the existing “tricks” such as deliberate mis-spelling. There are some29 who view the use of such tricks and “extraneous” information as being contrary to the letter (and even spirit) of the Turing, but for us it makes far more pragmatic sense to aim at passing the Turing with them, and then gradually stripping them away to (if possible) pass the Turing without them.
Second, the use of a virtual world enables several interesting variations on the explicit, one-on-one, Turing Test. For instance we could stage tests where:
The PC was undeclared and part of a group conversation (the Covert Group test) An PC was declared present (but unidentified) as part of a group conversation (the Overt Group test) The PC was undeclared and part of a set of one-on-one conversations (the Covert Singleton test) An PC was declared present (but unidentified) as part of a set of one-on-one conversations (the Overt Singleton test, the original Turing)
Even in our development of PCs? we are already encountering cases of the Covert Group test without setting them up, and with our PCs? running 24 hours a day then cases of Covert Singleton testing may be occurring every day.
Third, the original Turing Imitation Game was based on the ability of a tester to discriminate a person of one sex imitating a person of the other. Again the virtual world environment provides a new space within which to conduct such testing – and again it is happening on a daily basis within Second Life (where one survey found as many as 50% of men playing female characters, but only 12% of females playing males)7. 8 Future Research
This paper has aimed to provide an introduction to the synergy between the Turing Test and virtual worlds, and to summarise our areas of existing, mainly commercial, research.
Whilst the development of the chatbot/PC engine itself will remain an area of commercial R&D, we believe that there is considerable scope to undertake academic research around the relationship between the Turing Test and virtual worlds. This could be based around one (or more) of the following key questions:
1. AAAI: http://www.aaai.org/AITopics/html/overview.html 2. Active Worlds: http://www.activeworlds.com (2008) 3. AIML: http://www.alicebot.org/TR/2001/WD-aiml/ 4. AIRA: Private correspondence Burden/Edoja 5. Alice Superbot: http://www.alicebot.org/superbot.html 6. Alice: http://www.pandorabots.com/pandora/talk?botid=f5d922d97e345aa1 (2008) 7. Amdahl K.: http://kateamdahl.livejournal.com/27275.html 8. Ballantyne T.: Divergence. ISBN 055358930X. Spectra Books (2007) 9. Chatterbox Challenge: http://www.chatterboxchallenge.com/ (2008) 10. Da Silva: http://gwynethllewelyn.net/2007/06/20/the-20000-question-an-essay-by-extropia-dasilva/ (2007) 11. Discourse: http://www.daden.co.uk/pages/000384.html# 12. Ekman P.: "Basic Emotions". In T. Dalgleish and M. Power (Eds.). Handbook of Cognition and Emotion. John Wiley & Sons Ltd, Sussex, UK (1999) 13. Eliza: http://www-ai.ijs.si/eliza/eliza.html (2008) 14. George: http://www.jabberwacky.com/george 15. Goertzel B.: Artificial General Intelligence, Ben Goertzel & Cassio Pennachin (Eds), Springer, (2007) 16. Ikea: http://126.96.36.199/ikea-us/cgi-bin/ikea-us.cgi (2008) 17. Kaneva: http://www.kaneva.com (2008) 18. Krenn B., Pirker H.: Defining the Gesticon: Language and Gesture Coordination for Interacting Embodied Agents, in Proceedings of the AISB-2004 Symposium on Language, Speech and Gesture for Expressive Characters, University of Leeds, UK, pp.107-115, (2004) 19. Kurzweil: http://www.kurzweilai.net/meme/frame.html?m=1 20. Loebner: http://www.loebner.net (2008) 21. Maslow A. H.: A Theory of Human Motivation. Originally Published in Psychological Review, 50, 370-396.(1943) 22. Mimicbot: http://www.daden.co.uk/docs/Daden%20Newsletter%200707.pdf (2007) 23. Open Cyc: http://www.opencyc.org/ 24. Ortony A, Clore G. L., and Collins A.: The Structure of Emotions. Cambridge University Press, (1988) 25. Pennachin C.: Contemporary Approaches to Artificial General Intelligence, Cassio Pennachin & Ben Goertzel, 2007 in Artificial General Intelligence, Ben Goertzel & Cassio Pennachin (Eds), Springer, (2007) 26. Pirker H. and Krenn B.: D9c: Report on the assessment of existing markup languages for avatars, multimedia and multimodal systems on the WWW at http://www.ofai.at/research/nlu/NECA/RRL/index.html (2002) 27. Piwek P., Krenn B., Schröder M., Grice M., Baumann S., Pirker H., In Proceedings of the Workshop ``Embodied conversational agents - let's specify and evaluate them!'', held in conjunction with AAMAS-02, July 16 2002, Bologna, Italy (2002) 28. RDF: http://www.w3.org/RDF/ 29. Robitron: email@example.com in Robitron List 14 Nov 07 (2007) 30. Second Life: http://www.secondlife.com (2008) 31. Slater S.: Private correspondence (2008) 31. Topic Maps: http://www.topicmaps.org/ 32. Turing, A.M.: Computing machinery and intelligence. Mind, 59, 433-460 (1950) 33. Wang P.: http://nars.wang.googlepages.com/home (2007) 34. Wikipedia: http://en.wikipedia.org/wiki/Chatterbot (2008) 35. Word Net: http://wordnet.princeton.edu/ 36. World of Warcraft: http://www.worldofwarcraft.com/index.xml 37. YOMA: http://www.youronlinemarketingagency.com/