Advanced Synthetic Characters, Evil, and E ∗

We describe our approach to building advanced synthetic characters, within the paradigm of logic-based AI. Such characters don’t merely evoke beliefs that they have various mental properties; rather, they must actually have such properties. You might (e.g.) believe a standard synthetic character to be evil, but you would of course be wrong. An advanced synthetic character, however, can literally be evil, because it has the requisite desires, beliefs, and cognitive powers. Our approach is based on our RASCALS architecture, which uses simple logical systems (first-order ones) for low-level (perception & action) and mid-level cognition, and advanced logical systems (e.g., epistemic and deontic logics) for more abstract cognition. To focus our approach herein, we provide a glimpse of our attempt to bring to life one particular advanced synthetic character from the “dark side” — the evil character known simply as E. Building E entails that, among other things, we formulate an underlying logico-mathematical definition of evil, and that we manage to engineer both an appropriate presentation of E, and communication between E and humans. For presentation, which we only encapsulate here, we use several techniques, including muscle simulation in graphics hardware and approximation of subsurface scattering. For communication, we use our own new “proofbased” approach to Natural Language Generation (NLG). We provide an account of this approach. The Dearth of AI in AI There’s an unkind joke — which made the rounds (e.g.) at the Fall 2004 AAAI Fall Symposium on Human-Level AI — about the need to create, within AI, a special interest group called ‘AI’. This kind of cynicism springs from the not uncommon, and not totally inaccurate, perception that most of AI research is aimed at exceedingly narrow problems light years away from the cognitive capacities that distinguish human persons.1 ∗The R&D described in this paper has been supported in part by much appreciated grants from AFRL-Rome and DARPA-IPTO. An endless source of confirming examples can be found in the pages of the Machine Learning journal. The dominant learning technique that you yourself employ in striving to learn is reading; witness what you’re doing at the moment. Yet, a vanishingly small amount of R&D on learning is devoted to getting a computer program to learn by reading. Human-level AI is now so unusual that an entire upcoming issue of AI Magazine will be devoted to the subject — a bit odd, given that, at least when the field was young, AI’s journal of record would have routinely carried papers on mechanizing aspects of human-level cognition. Seminal AI thinkers like Simon, Newell, Turing — these researchers didn’t shy away from fighting to capture human-level intelligence in machine terms. But now their attitude seems moribund. But gaming, simulation, and digital entertainment (and hereafter we refer simply to ‘gaming’ to cover this entire field/market), thankfully, are different: ultimately anyway, they call for at least the appearance of human-level AI (Bringsjord 2001). (On a case-by-case basis, as various games show (e.g., The Sims (Electronic Arts Inc. 2000)), a non-advanced character will of course do just fine.) Gaming doesn’t strive just for a better SAT-based planner, or another tweak in a learning algorithm that doesn’t relate in the least to human learning. A SAT planner doesn’t constitute a virtual person. But that’s precisely what we want in gaming, at least ultimately. And even in the short term we want characters that at least seem human. Methodologically speaking, gaming’s best bet for characters that seem human is to bite the bullet and strive to engineer characters that have what it takes to be human. This, at least, is our strategy. Gaming and Full-Blown Personhood Now, there are various ways to get clearer about what gaming, at least in the long-term, needs when it comes to humanlevel intelligence. One way is to say simply that gaming needs artificial creatures which, behaviorally at any rate, satisfy one or more plausible proposed definitions of personhood in the literature. One such definition has been proposed by Bringsjord in (Bringsjord 1997). This definition essentially amounts to the view that x is a person if and only if x has the capacity 1. to “will,” to make choices and decisions, set plans and projects — autonomously; 2. for consciousness, for experiencing pain and sorrow and happiness, and a thousand other emotions — love, passion, gratitude, and so on; 3. for self-consciousness, for being aware of his/her states of mind, inclinations, preferences, etc., and for grasping the concept of him/herself; 4. to communicate through a language; 5. to know things and believe things, and to believe things about what others believe, and to believe things about what others believe about one’s beliefs (and so on); 6. to desire not only particular objects and events, but also changes in his or her character; 7. to reason (for example, in the fashion exhibited in the writing and reading of this very paper). Unfortunately, this list is daunting, especially if, like us, you really and truly want to engineer a virtual person in the short term. A large part of the problem is consciousness, which we still don’t know how to represent in thirdperson machine terms (Bringsjord 1998; Bringsjord 2001). But even if we leave aside consciousness, the rest of the attributes in the above list make for mighty tough challenges. In the section “Making the Challenge of Personhood Tractable” we shall retreat from this list to someting doable in the near term, guided by particular scenarios that make natural homes for E. But in the end, whatever appears on this list is an engineering target for us; in the long term we must confront each clause. Accordingly, in the section “How Does E Talk?” we explain how we are shooting for clause 4, communication. We have made progress on some of the other clauses, but there is insufficient space to present that progress herein. Clause 5 is one we believe we have pretty much satisfied, via the formalization and implementation given in (Arkoudas & Bringsjord 2005).2 Current State of the Art versus Computational Persons Synthetic Characters in Gaming What’s being done now in gaming, relative to full-blown personhood, is clearly inadequate; this can be quickly seen by turning to some standard work: Figure 1 shows an array of synthetic characters from the gaming domain; these will be familiar to many readers.3 None of these creatures has anything close to the distinguishing features of personhood. Sustained treatments of synthetic characters and how to build them are similarly limited. For example, consider Figure 2, taken from (Champandard 2003).4 As a mere FSA, there is no knowledge and belief, no reasoning, no declarative memories, and no linguistic capacity. In short, and this is perhaps a better way of A preprint is available online at http://kryten.mm.rpi.edu/arkoudas.bringsjord.clima.crc.pdf. Worst to best, in our eyes: Top-left, The Legend of Zelda; SC spits text upon entering room. Top-right, Chrono Trigger; treebranching conversations. Middle-left, Might & Magic VI (Shopkeepers). Middle-right, Superfly Johnson from Daikatana; behavior scripting, attempts to follow player and act as a sidekick (fails!). Bottom-left, Galatea – Interactive Fiction award winner for Best NPC of 2000 (text-based). Bottom-right, Sims 2. But even here, nothing like what our RASCALS architecture has is present. This is an excellent book, and it’s used in our lab for building synthetic characters. But relative to the loftier goals of reaching bona fide personhood in artificial characters, there’s clearly a lot of work to be done. putting the overall problem infecting todays’s virtual characters, all of the cognitive capacities that distinguish human persons, according to the science of cognition (e.g., (Goldstein 2005)), are missing. Even the state of the art using cognitive architectures (e.g., SOAR) is primitive when stacked against full-blown personhood (Ritter et al. June 2002). Figure 1: Sample Synthetic Characters What About Synthetic Characters in Cutting Edge Research? What about research-grade work on synthetic characters? Many researchers are working on synthetic characters, and have produced some truly impressive systems. However, all such systems, however much they appear to be human persons, aren’t. We now consider three examples of such work, and show in each that the character architectures don’t have the underlying cognitive content that is necessary for personhood. REA An agent developed by (Cassell et al. 1999) known as REA is an example of a successful, robust agent whose developers focused primarily on embodied conversation and the conversational interface. She is described as being an expert in the domain of real estate, and interactions with REA are both believable and informative. REA, however, is representative of many of the industry’s most successful agents in that she excels at content management, but fails to deliver rich emotive and cognitive functionality. REA, after all, cannot generate English from arbitrary underlying knowledge. Like many of her peers, REA’s underlying cognitive capabilities are modeled in an ad-hoc fashion. Her personality is in no way defined; her interactions within a particular situation lack subtlety and depth. While she excels as a simulated character and a conversational agent, she is bereft of the rich cognitive content with which advanced synthetic characters must