Greg Detre
Tuesday, March 11, 2003
The heart of the AIBO seems to be the ethogram (categorisation of behavioural patterns), based on ethological observation of canines, that provides the top level of behavioural subsystems. Arkin et al. list the following:
investigative (searching/seeking)
sexual
epimeletic (care and attention-giving)
eliminative (excretion and urination)
et-epimeletic (attention-getting or care-soliciting)
ingestive (food and liquids)
allelomimetic (doing what others in group do)
comfort-seeking (shelter-seeking)
agonistic (associated with conflict)
miscellaneous motor
play
maladaptive
They state that the �play� and �maladaptive� subsystems are treated as separate behavioural subsystems for pragmatic reasons, which makes sense. One would expect that �maladaptive� behaviour has been added in order to group together any idiosyncracies of canine behaviour that the simplified AIBO model fails to produce. In a similar way, we might imagine that playful behaviour could arise to some degree (e.g. as a mechanism to facilitate learning) within most of the other subsystems, but merits emphasising because of the AIBO�s function as a toy.
Interestingly, the particular subsystems in the ethogram provided mostly consist of motivational and emotional states, like being curious, affectionate, anxious, aggressive, hungry etc. This chimes well with the growing perception in cognitive psychology that emotions are cognitive modes that flexibly influence behaviour (e.g. Minsky, �The emotion machine�; Rolls, �The Brain and emotion�).
Each subsystem can be broken down further into a tree of modes, then even lower-level stimulus-response modules. The article implies that the AIBO is entirely deterministic in its choice of actions, using lateral inhibition (i.e. there is no small stochastic element that could make its behaviour unpredictable). In order to choose between them, actions are first evaluated according to a �homeostasis regulation rule�, which attempts to satisfy the system�s internal drives (hunger, thirst, elimination, tiredness, curiosity and affection) especially when the appropriate external stimuli are available. The AIBO continually occupies a point in a three-dimensional emotional space that is a function of this internal regulatory state and the presence of external stimuli. If the internal variables are within the regulatory range, the pleasantness is high. Circadian rhythm and unexpected stimuli affect arousal. Confidence relates to the certainty of recognised external stimuli. Together, these three dimensions can locate Ekmann�s six basic emotional states (happiness, anger, sadness, fear, surprise and disgust).
Probably of more interest than the ethological aspect of the architecture is the AIBO�s use of emotionally-grounded symbols, as a means of enriching its conception of objects, by �[learning] the meanings of the acquired symbols in terms of the robot�s needs�. Arkin et al. rather simplistically equate the �meaning� of a symbol with the �correct behavioural response�. This approach turns out have some powerful implications though � by incorporating knowledge of the change of internal state elicited by a given behaviour response into each symbol, AIBO can learn how to use its environment to systematically affect its internal state. I can see how emotionally grounding symbols leads to a far richer conception than simply associating physical (i.e. visual and auditory) information, and presumably this system works because it associates not just the current emotional state, but also the resulting change in emotional state. In the case of the SDR-4X, the authors argue that �the emotionally grounded concept is a key to understanding the meaning of the user�s uttered words in relation to the robot�s perceptions, behaviours, capabilities and needs�. However, their use of emotionally-grounded symbols still stretches �meaning�. I take Minsky�s view that the richness of a concept�s �meaning� is in proportion to its multiplicity of representation, and the AIBO�s is still pretty impoverished.
Infant rats suckle as their only means of nutrition for roughly their first 20 days. Around this time, they start to ingest liquids and semi-solids. The standard view, that Hall & Williams successfully debunk, is that the suckling and ingestive behaviours are developmentally continuous. The assumption of �developmental continuity� is a tempting (and parsimonious-seeming) one to make when two behaviours serve a similar function, utilise similar body parts, and don�t appear to co-exist. As Hall & Williams put it, �the system examined in early development is viewed as the infantile behavioural form that is gradually elaborated into the adult pattern�, which often amounts to saying that the two behaviours share a neural substrate that changes gradually over time.
The suckling behaviour of infant rats is highly-stereotyped and present even in poorly-developed and motorically imature newborn pups. Moreover, depriving pups of suckling stimulation has no effect on the initiation of suckling until after 10 days of age, and for the first two weeks pups continue to suckle whether or not they get milk from the nipple. Indeed, pups of less than two weeks showed little inclination to control intake according to nutritional state by stopping suckling, even when nearly drowning themselves. Taken together, these facts indicate that suckling is a highly prepotent behaviour.
They discuss five major ways in which ingestive and suckling behaviour can be distinguished:
1. Response differences
Adult rats lick, lap, mouth and chew their food in contrast to the unique and stereotypic stretch response during suckling
2. Differences in the external control of behaviour
Odour cues are important for suckling, but not for feeding. More importantly, the initiation of suckling is not affected by nutritional deprivation, as feeding is.
3. Differences in the internal control of behaviour
Internal physiological controls (e.g. gastric distension, satiety hormones) hardly affect the initiation of suckling (although the issue is complicated by infant rats� ability to partially control rate of intake).
4. Differences in neural substrates
Various pharmacological studies showing differential effects of a drug on the two behaviours, e.g. amphetamines depress adult ingestion but increase suckling, and 5-HT receptors appear to instil a preference for either suckling or ingestion, even much earlier/later than usually found. However, lesion studies have not so far supported the idea of separate neural substrates.
5. Differential experiential determinants
While suckling is highly vulnerable to disruption by suckling deprivation, feeding is not.
Hall & Williams produce evidence that the ingestive system co-exists with the suckling system, though remains dormant for the first 2-3 weeks of the pup�s life, and can be coaxed in pups separated from their mothers. Even 1-day old deprived pups actively licked and swallowed milk infused into the front of their mouths, and even newborn pups consumed significant volumes of milk, sucrose or wet Purina Chow mash spread on towels on the floor of test containers (in warm conditions). These feeding responses, rather than eliciting the stereotyped stretching response associated with suckling, led the pups to unexpectedly produce an enormous range of behavioural fragments, like mouthing, licking, crawling, probing and tumbling. Furthermore, unlike with suckling, the intake volumes when feeding in this manner were proportional to deprivation.
They characterise pups� independent ingestion as �a hidden forerunner to later ingestion that will only reveal its presence if tricked in the right manner�. In contrast, suckling is a special-purpose module with a preset time-frame, or what Oppenheim terms a �transient ontogenetic adaptation�.
Hall &
Williams� conclusions seem pretty firmly supported, and should direct attention
to alternative �transient ontogenetic adaptations�, and make us wary of
assuming continuity between different behaviours that serve similar functions
during a system�s development. They point towards human infants� automatic
grasp reflex of the hand and foot, coordinated stepping pattern and rooting
suckling reflexes of the mouth, all of which become increasingly difficult to
elicit, but which may reappear with senility and brain damage. This suggests
the possibility for further confirmation of Hall & Williams� claim � if they
could demonstrate that lesioning adult rats could damage the feeding response
but re-elicit the suckling response, I would certainly have few remaining
doubts.
My only
final query related to the discussion of neural substrates. It seems possible
to me that two systems could be developmentally discontinuous, coexist, be
elicited by different stimuli and result in very different behaviours, and
serve similar functions at different times in the animal�s development (like
suckling and feeding), but still share a neural substrate. Hall & Williams
cite little anatomical, recording or imaging evidence to support the
pharmacological evidence for the two systems having separate substrates.
Indeed, the evidence from lesions affected neither or both systems. One possible
explanation for this could be that the two systems have an (at least partly)
shared substrate, but are opposing attractors. Even if this were the case, we
would still be able to talk of two separate systems, simply occupying the same
substrate (like different programs running on the same computer).