Simulations in R to explore the evolutionary advantage of an ambiguity processor distinct from and often competing with a risk processor based on frequencies.
Christian Luhmann wrote on 23 July 2012
9:54 AM (1 hour ago)
I was thinking about the snake
simulation. Right now, if I correctly understand the simulations, the
agents (the foragers) face snakes with fixed nutritional value and
variable (binary) status (poisonous vs. not). The agents then
probabilistically choose to eat feed or flee. In the simulations you
sent me that's about all there is, but I recall Jack discussing a
version where some agents included a probabilistic representation of
their uncertainty about the poison/not state of snakes a other agents
worst-case representation. However, my understanding was that the
agents uncertainty was always about this boolean poisonous/not
proposition. (Again, this is all pieced together from very old
discussions, so I may be off).
I was thinking about the case in which the uncertainty was instead
about a continuous variable (e.g., how poisonous a snake is) and in
which actions (feed or flee) was a function of expectations of those
actions (e.g., expectation about how poisonous snakes are and the
agent's current level of fitness). Would that "fix" the time scale
issue? Agents with probabilistic uncertainty would need full PDFs
rather than single probabilities. When encountering extreme but rare
values (the one super-poisonous snake), the probabilistic agents would
note such extreme values as possible but unlikely whereas the
worst-case agents would essentially "forget" all less-poisonous
snakes. If the world suddenly and unexpectedly got much, much worse,
the worst-case agents would have a chance to cope whereas the
probabilistic agents would try to learn about the new, shifting
distribution of poison level, but presumably die in the process.
The basic difference picks up on something that we have run into in
other contexts before: single observations of a discrete/binary
variable provides very little information about the probabilistic
distribution over those discrete/binary values (particularly about the
tails). For example, observing 1 poisonous snake tells you that
P(poisonous) cannot be 0, but not a whole lot more than that. In
contrast, observing a highly poisonous snake tells you that such
things exist and, if that's all you care about, that may be enough.
Like the loss aversion story, as long as things are going well, such
overly conservative attitudes may be beneficial.
Anyway, I'm not sure that makes a lot of sense. It may be the kind of
thing that is more easily explained in person, but I thought I would
write it down while I still remembered it. If I have some time
(unlikely), I may try to code up a little simulation to at least
formalize things.
C
Jack Siegrist
wrote on 23 July 2012 10:07 AM (1 hour ago)
Maybe the snake business could be seen as a matter of the advantages of instinct vs learning in different situations. There must be some biological theory about when it is better to decide on instinct vs decide on experience. Maybe I will try to see if there is anything relevant.
Jack