Here are the slides from my position paper at the AAAI Fall Symposium on Naturally Inspired Artificial Intelligence.

I’ve just put up the final version of my position paper for the AAAI Symposium on Intelligent Narrative Technologies.  It argues that popular narratives deal at least as much with the ways characters depart from ideal rationality as it deals with actual rational behavior, that we don’t have computational theories of those modes of behavior, and that we could learn a lot by trying to model them.

I was working yesterday on adding fighting behavior to my procedural character animation.  In particular, I wanted the kids to be able to shove one another around.  But due to a bug they kept approaching one another rather than stopping at “shoving distance.”  So they grapple in what looks like a particularly comical make-out session.  Possibly useful, but best saved for the parent characters…

I’ve been trying to think about anxiety recently.  I want to argue that anxiety is part of a broader meta-level control system involved in risk assessment.  The obvious scientific response to this would be something along the lines of “well, duh.”  But I’ve had a great deal of difficulty articulating exactly what I mean to my friends.  So I’m writing this overly long post to try to work through it.  I’ll try using two different analogies, but first I’ll try to give some motivating background.

Why it isn’t obvious

When I started out working on my current project, I intended to include a high-level simulation of Gray and McNaughton’s revised (2000) version of Reinforcement Sensitivity Theory.  RST is a neuropsychological theory of anxiety that has been influential in the personality theory community.  RST argues (in the revised version) that there are three main behavioral subsystems:

• The Behavioral Activation System (BAS), which is responsible for approach behaviors,
• The Fight/Flight/Freezing system (FFFS), which is responsible for avoidance behaviors, and
• The Behavioral Inhibition System (BIS), which resolves conflicts between the two (or within one system).

In RST:

• Anxiety is triggered by goal conflicts detected by the BIS.
• Anxiety is responded to by the BIS (1) directing attention toward the source of the conflict to gather more information, then (2) resolving the goal conflict by assessing the stimulus and inhibiting one option while stimulating another.  The system has a bias toward the FFFS, so it’s more likely to resolve the conflict through some kind of escape behavior than through approach.

Along the way, I got distracted by the Attachment Theory literature.  In attachment experiments:

• Anxiety is triggered by the non-availability (physical or emotional) of the care-giver, unfamiliar environments, and the presence of strangers.
• Anxiety is responded to by activation of the Attachment Behavior System, which causes the child to try to look to the caregiver, achieve eye contact with them, or get physically closer to them, depending on the level of anxiety.
• (Note however, that Attachment Theory doesn’t claim these are the sole causes or responses to anxiety).

A securely attached child need not maintain proximity to the parent at all times, but uses the parent as a “secure home base” from which to make exploratory excursions.  Essentially, the child hangs out with the care-giver, charges up his/her security battery, then goes out and explores until the security battery gets depleted, then recharges, either by actually returning to the caregiver, by achieving eye contact with him/her, or simply by making sure the child still have line-of-sight to the care-giver.

The thing to notice here is that RST models anxiety as being by definition goal conflict.  Attachment theory doesn’t define anxiety but includes among the things that can trigger it, the inaccessibility of the care-giver.  But inaccessibility isn’t goal conflict.  It’s closer to goal failure, although the fact that child’s desire for parental availability waxes and wanes means it’s more complicated than a nice simple maintenance goal like Always(Near(parent)).

In other words, the neural basis for anxiety provided by RST probably isn’t the neural basis for what attachment theory calls anxiety.  So the question then is whether the phenomenon RST is calling anxiety and the one attachment theorists call anxiety are the same phenomenon, and if so, what the phenomenon “really” is.

An immune system analogy

Your immune system can be grossly divided into the specific and non-specific immune systems, (or adaptive and innate).  Each has a set of detection mechanisms and a set of response mechanisms.  The specific immune system detects threats using antibodies that recognize specific proteins on the threats and responds by (among other things) attacking whatever is attached to the naughty protein.  The non-specific system notices general signs that something is awry, such as chemical markers of (non-programmed) cell death, and responds with general strategies to discourage infection, such as fever and inflammation, as well as general sensitization of the rest of the immune system.

The fight/flight system is also essentially a system for dealing with specific threats; fight and flight are only possible when a specific aversive object has been identified to attack or avoid.  But there are a number of stimuli that aren’t threats themselves, but rather general cues that you might be in a more dangerous situation than you’re used to.  Examples of these would be:

• Darkness
• Unfamiliar environments, situations or people
• Losing contact with a care-giver, or more generally, being alone
• Displays of anxiety in those around you
• The world behaving in ways it isn’t supposed to

Consequently, your only possible responses are to:

• Activate safety-seeking behaviors, such as attachment, social engagement (the “tend and befriend” system), or withdrawl (hiding)
• Increase your vigilance level
• Adopt a risk-averse action-selection strategy, e.g. by biasing action selection toward the FFFS
• Prepare for any concrete threats that might appear, e.g. by releasing adrenaline

Notice that both the stimuli and responses listed above are generally associated with anxiety.

One can argue that this works well for the kinds of anxiogenic stimuli considered in attachment theory, but it doesn’t work well for goal conflict.  But while goal conflict isn’t a cue to an increased risk of encountering threatening objects, it is a cue that you’re moving outside of situations you know how to handle easily; the BIS doesn’t fire for all goal conflicts, just goal conflicts you haven’t already learned to resolve.  So again, it’s a cue to possible future problems and so argues for shifting to a more conservative strategy.

a Vygotskian analogy

Vygotsky was a developmental psychologist who, among other things, developed the notion of the zone of proximal development.  The idea is that when you’re learning a domain, the tasks within it fall into those you’ve already mastered, those that are too damn hard, and those you don’t know yet but are ready to learn.  The latter group forms the zone of proximal development.  An instructor’s job is to keep the student in the ZPD by taking responsibility for the parts of the task that are currently too hard for the student to learn, a process known as scaffolding.   As the student learns more, the ZPD shifts, and the instructor can gradually remove the scaffolding.  (I’m a fan of this model in part because I think the reason programming is so hard to teach is that programming is very difficult to scaffold effectively).

The important concepts are that there’s:

• A continuum of activity, from too hard to too easy, and
• A set of scaffolding techniques that can be modulated to steer the learner toward a sweet spot in the continuum

However, there’s no reason the scaffolding need be done exclusively by a teacher.  To the extent that it’s possible to make the world seem harder or easier (at a cost) by shifting between risk seeking and risk aversion, or by controlling the level of dependence on a care-giver, we can think of the child as self-scaffolding.  In fact, we can even think of adults as self-scaffolding to the extent that they modulate these behaviors, even if they aren’t necessarily learning.  So we can think of anxiety as the feedback signal that’s telling the agent it needs to increase its level of scaffolding.

So in other words, the argument is that there’s a general meta-level process of security assessment that modulates both high-level parameters like risk aversion, and remediation behaviors like attachment.  It’s essentially a homeostatic process.  You want enough risk aversion and remediation to attain an acceptable level of security, but not so much that wasting resources preparing for threats that probably aren’t there.  Anxiety is the signal to increase the response of the security system, while other signals like boredom would tend to decrease its activity.

There.  Does that make sense?

Technorati Tags: cognitive architecture, social behavior, neurophysiology, ai

Robot hackers have long realized that human observers tend to over-attribute intelligence, or at least intentionality, to robots, at least provided that they’re moving.  (Dave) Miller’s law states that the perceived intelligence of a robot is directly proportional to its velocity (Dave didn’t name this Miller’s law, but he said it once at a workshop and I’m fond of quoting it).

The image above is a screenshot of what appears to be two child characters playing with one another while being watched by an adult.  However, in actuality, what I’ve been implementing is attachment behavior, which is the response of children to stress by seeking out their caregiver (more on this another day).  To implement that, I need to have something to stress the kids out.  The right way to do it is to implement a real social engagement system with wariness and coy behaviors, play, turn-taking, etc.  However, the first step in that is simply to make a second child and then hack the children’s appraisal systems to assign negative valence to strangers (i.e. to each other).  All that does is make the kids watch one another and keep their distance from one another.  For example, one won’t approach the ball if the other is too close to it.  There’s no real sociality going on there.

The interesting thing is that it’s enough to make them look like they’re playing.  They both run to the ball, but then when one gets to close to it, the other backs off.  The first one will kick it until it happens to kick it toward the other one (which is pretty frequent since I haven’t implemented aiming).  At that point, the first one stays away from the ball and the second one plays with it.  This continues until they get far enough from the parent to engage the attachment system, at which point the attached child runs to the parent and hugs him/her, then runs back to play.

The point of this isn’t that this is a good simulation of anything, just that surprisingly simple behavior can appear engaging and intelligent, provided that whatever behavior you do have is relatively fluent.

I just finished reading a paper by Stephen Porges on the Polyvagal theory.  It’s very cool.  Here’s my attempt as a non-specialist to summarize it.  The traditional view of the autonomic nervous system is that it has two opposing processes, the sympathetic and the parasympathetic systems, which speed up and slow down the heart, respectively, as well as shifting metabolic resources between organs.  The sympathetic system, which is stimulated in part by the perception of threat, raises heart rate and metabolism, and shifts energy to the brain and muscles, thus preparing the body for a fight-or-flight respond to the perceived thread.  The parasympathetic system acts through the vagus nerve to slow the heart rate and/or shift resources to the gut to aid in digestion.

Porges makes two arguments about the vagal (parasympathetic) system:

• The vagal system isn’t just used for reallocating energy resources during rest and digestion, but it’s also part of the neural substrate for the freeze response (a different, and in fact phylogenetically older, threat response).
• In mammals, the vagal system is split into two systems – the older dorsal system and a newer ventral system.  This newer system is also connected to facials, vocal, and neck muscles, and forms part of the social engagement system, a newer threat response system that’s only present in (social) mammals.

He argues that the autonomic system’s threat responses are organized roughly as a Brooksian subsumption hierarchy, with phylogenetically newer systems overriding older systems in when they conflict.  This is interesting if you’re interested in biologically-based AI because it gives us a much better understanding of the low-level threat-response systems and also shows that social behavior is wired in at surprisingly low levels.

Porges also makes two claims that as far as I can tell are logically independent of the poly-vagal theory, but which are nonetheless interesting.  One is that the (rare) phenomenon of voodoo death, in which people are literally scared to death, is not actually due to overactivation of the sympathetic nervous system overdriving the heart, but rather to the overactivation of the older parasympathetic (in this case, the dorsal system) slowing the heart down to the point where it can no longer supply enough oxygen for itself, much less the brain.  This apparently better matches the experimental results in a rather horrible rat experiment that was done in the 50s.  The claim is that it’s a case of a response that was adaptive for older species being carried over to mammals, in whom it is fatal because of their higher metabolic requirements.

The other claim is based on the observation that there are muscles in the inner ear, which modulate its mechanical properties so as to accentuate or attenuate certain frequencies, in particular, the frequencies associated with the human vocal tract.  He proposes that in at least some cases of developmental disorders involving lack of social engagement, it could be because of understimulation of these muscles, making it harder for children to attend to human voice, and thereby leading to understimulation of the social engagement system.  The description of the experiment in the paper is vague and I haven’t followed up to read the real papers on it, but he found that when he played computer-generated sounds that emphasized frequencies in the range of human speech, most of the children showed “noticable improvements in social behavior and communication skills following the intervention.”  Amazing.