Re: about modeling human decisions?
From: Glen M. Sizemore (gmsizemore2_at_yahoo.com)
Date: 03/11/05
- Next message: robert j. kolker: "Re: Epistemology 201: The Science of Science"
- Previous message: Bruce Weaver: "Re: about modeling human decisions?"
- In reply to: Stephen Harris: "Re: about modeling human decisions?"
- Next in thread: Wolf Kirchmeir: "Re: about modeling human decisions?"
- Messages sorted by: [ date ] [ thread ]
Date: Fri, 11 Mar 2005 07:52:06 -0500
"Stephen Harris" <cyberguard1048-usenet@yahoo.com> wrote in message
news:tV9Yd.17088$Pz7.13660@newssvr13.news.prodigy.com...
>
> "Bob Wheeler" <bwheeler@echip.com> wrote in message
> news:cc11c$4230db9b$466ea752$9522@nf1.news-service.com...
> > Stephen Harris wrote:
> >> "Bob Wheeler" <bwheeler@echip.com> wrote in message
> >> news:9aae5$42305f54$466ea752$12937@nf1.news-service.com...
> >>
> >>>Bruce Weaver wrote:
> >>>
> >>>>David Jones wrote:
> >>>>
> >>>>
> >>>>>kiki wrote:
> >>>>>
> >>>>>
> >>>>>>HI all,
> >>>>>>
> >>>>>>I want to create self-automata models of human beings and model how
> >>>>>>human-beings collectively make decisions and how people with
> >>>>>
> >>>>>
> >>>>>different
> >>>>>
> >>>>>
> >>>>>>levels of knowledge and sense influence each other...
> >>>>>>
> >>>>>>Are there any researches about how to build this kind of models?
> >>>>>>
> >>>>>>Thanks a lot!
> >>>
> >>>>Here's another paper to look at:
> >>>>
> >>>
> >>>To take another tack, the "rational man" idea assumes that humans have
> >>>infinite calculating capability and that game theory, for example,
> >>>represents actual human behavior. In fact, humans have very limited
> >>>calculating ability. They solve problems by heuristics, which are
simple
> >>>rules quite consistent with the actual functioning of the brain. For
> >>>example, in choosing the "best" of several items, humans will tend to
> >>>pick the first, or last one, or the one which is most familiar, and
will
> >>>ignore almost all of the "technical" characteristics of the items.
> >>>
> >>
> >>
> >> "The frameworks of game theory and mechanism design have exerted
> >> significant influence on formal models of multiagent systems by
> >> providing tools for designing and analyzing systems in order to
> >> guarantee certain desirable outcomes. However, many game theoretic
> >> models assume idealized rational decision makers interacting in
> >> prescribed ways. In particular, the models often ignore the fact
> >> that in many multiagent systems, the agents are not fully rational.
> >> Instead, they are computational agents who have time and cost
> >> constraints that hinder them from both optimally determining their
> >> utilities from the game and determining which strategies are best
> >> to follow. ...
> >> http://www.cs.uwaterloo.ca/~klarson/papers/LarsonThesis.pdf
> >>
> >> Because of this, the game theoretic equilibrium for rational agents
> >> does not generally remain the same for agents with bounds on their
> >> computational capabilities. This creates a potentially hazardous
> >> gap in game theory and automated negotiation since computationally
> >> bounded agents are not motivated to behave in the desired way.
> >>
> >> My thesis statement is that it is possible to bridge this gap. By
> >> incorporating computational actions into the strategies of agents, I
> >> provide a theory of interaction for self-interested computationally
> >> bounded agents. This allows one to formally study the impact that
> >> bounded rationality has on agents' strategic behavior. It also
> >> provides a foundation for game-theory and mechanism design for
> >> computationally limited agents."
> >
> >
> > Interesting. It will take a while for me to read this. How does this
> > limitation on computation square up with actual psychological studies
that
> > suggest the "human mechanism" is comparative rather than computational?
> >
> > --
> > Bob Wheeler --- http://www.bobwheeler.com/
>
> SH: I'm not sure, it seems debatable still.
>
> http://spartan.ac.brocku.ca/~lward/Morgan/Morgan_1903/Morgan_1903_16.html
> "As the outcome of a very careful consideration of the whole question,
> supplemented by a number of interesting and valuable experimental
> observations, he [L. T. Hobhouse in "Mind in Evolution"] concludes
> that "the highest animals have as much capacity for dealing with the
> practical exigencies of their surroundings as can be attained by an
> intelligence limited in its scope to the concrete and the practical.
> Intelligence as we conceive it in this stage is capable of forming
> what we have called practical judgments."
>
> If then behaviour which is the outcome of concrete sense-experience,
> is placed in the game category as rational conduct based on the
> conceptual thought which results from the analysis of experience and
> the synthesis of ideal construction, we must freely admit that animals
> can and do reason. But I have used the term reason in a more restricted
> sense. Mr Hobhouse regards such restriction as arbitrary. The whole
> question, in his view, is a matter of degree. " It is not that new
> faculties are introduced, but that old faculties receive a fresh
> development." [SH: ... "practical judgments as contrasted with the
> universal judgments of conceptual thought."]
> -------------------------------------------------------------------
>
> I'm going to quote an important part of Kate Larson's thesis:
>
> http://www.cs.uwaterloo.ca/~klarson/papers/LarsonThesis.pdf
> 8.3 Sensitive Mechanisms (page 177)
> "In this section we study sensitive mechanisms in order to
> understand whether it is possible to design mechanisms which
> are non-deliberative, deliberation-proof and non-deceiving.
> We start by studying the properties of direct mechanisms.
>
> Theorem 27. There exists no value-based sensitive direct
> mechanism that is deliberation proof across all problem
> instances. (An instance is dedined by the agents performance
> profiles, cost functions, and current problem instance.)
>
> Theorem 28. There exists no sensitive value-based mechanism that is
> * non-deliberative,
> * deliberation-proof, and
> * non-deceiving
> across all problem instances. (An instance is defined by agents'
> performance profiles, cost functions).
> ...
> 8.4 Summary
> In this chapter we laid out mechanism design principles for
> computationally limited agents. We first showed that the revelation
> principle applies to such settings in a trivial sense by having the
> mechanism carry out all the computing for the agents. This is
> impractical, and we proposed that mechanisms should be
> non-deliberative: the mechanism should not be solving the deliberation
> problems for the agents. Second, mechanisms should be deliberation-proof
> : agents should not deliberate on others' valuations in equilibrium.
> Third, the mechanism should be non-deceiving: agents do not
> strategically misrepresent. Finally, the mechanism should be sensitive:
> the agents' actions should affect the outcome. We showed that no
> direct-revelation mechanism satisfies these four intuitively desirable
> weak properties. Moving beyond direct-revelation mechanisms, we showed
> that no value-based mechanism (that is, mechanism where the agents are
> only asked to report valuations - either partially or fully determined
> ones) satisfies these four properties.
>
> This result is negative. It states that either we must have mechanisms
> which do the computing for the agents, or complex strategic (and costly)
> counterspeculation can occur in equilibrium. However, there is some
> hope. It may be possible to weaken one of the properties slightly, while
> still achieving the others. For example, it may be possible to design
> multi-stage mechanisms that are not value based; the mechanism could
> help each agent decide when to hold off on computing during the mechanism
> (and when to compute on one's own valuation for different bundles of
> items in a combinatorial auction). In another direction, by relaxing
> strategic deliberation and compensating agents appropriately, it may
> be possible to design mechanisms where agents who can solve problems
> cheaply and efficiently do so for all agents."
>
> SH: Now from another paper about Nash equilibria and undecidability:
> http://garnet.acns.fsu.edu/~kprasad/toff.pdf
> "The Rationality / Computability Trade-off in Finite Games"
> Abstract:
> "The computability of Nash equilibrium points of finite games
> is examined. When payoffs are computable there always exists an
> equilibrium in which all players use computable strategies. However,
> there is a computable sequence of games for which the equilibrium points
> do not constitute a computable sequence. For this reason, there can be
> no algorithm that, given arbitrary payoffs, computes a Nash equilibrium
> point for the game. Even for games with computable equilibrium points,
> best responses of the players may not be computable. In contrast,
> approximate equilibria, and error-prone responses are computable." ...
>
> "Nash's theorem states that any finite game will have at least one
> equilibrium point. This paper starts by showing that if payoffs are
> computable then there is also an equilibrium in which each player
> uses computable strategies. In contrast, once we leave the framework
> of finite games it is possible to construct examples of games
> (Specker, 1959) for which the only equilibrium strategies are
> uncomputable real numbers. A logically separate question is how this
> equilibrium is reached. It is typically asserted that some form of
> evolution, learning, or introspection leads to an equilibrium being
> played. The computability literature argues that a minimal requirement
> for attainability of equilibrium is that there exist some algorithm
> that, given the complete description of a game, computes an equilibrium
> point.
>
> As it turns out, there cannot exist an algorithm capable of computing
> an equilibrium point for every instance of a game with computable
> payoffs. This negative result is a consequence of the fact that there
> exists a computable sequence of games for which the equilibrium points
> do not constitute a computable sequence. The result on the failure of
> computability of best responses for sequences shows that an individual's
> strategy in equilibrium cannot be constructed from the equilibrium
> strategies of other players either.
>
> If we require strategies to be computable from the full description
> of the game, these results present the following trade-off. One may
> either assume that agents are content with near-optimal decisions
> (and play an epsilon-equilibrium), or else assume they use procedures
> that are optimal for only a subset of the possible games they might
> encounter. In a somewhat different approach, individuals may be prone
> to error when making decisions.
>
> A number of bounded rationality models (e.g. Rosenthal, 1989, McKelvey
> and Palfrey, 1996) proceed by assuming that players play suboptimal
> strategies with small probability. In contrast to best-response play, such
> error prone responses will often be computable in terms of the strategies
of
> other players.
>
> We see that computability of best responses may fail for sequences
> (even when equilibria are computable from payoffs). Strikingly, if
> one allows for less than optimal (or error-prone) responses positive
> results become available."
>
> SH: It is difficult to place people in identical situations and then
> determine
> their actual behavior.
Yeah - it is called "doing science."
>Instead, people are given surveys with hypothetical
> questions about what would they do in some situation. These people then
> predict what they think they would do, in order to answer the survey.
Yeah - it is called "NOT doing science."
>
> Regards,
> Stephen
>
>
>
>
- Next message: robert j. kolker: "Re: Epistemology 201: The Science of Science"
- Previous message: Bruce Weaver: "Re: about modeling human decisions?"
- In reply to: Stephen Harris: "Re: about modeling human decisions?"
- Next in thread: Wolf Kirchmeir: "Re: about modeling human decisions?"
- Messages sorted by: [ date ] [ thread ]
Relevant Pages
|
