Introduction to Human Heuristics
3 likes2,638 views
The document discusses human heuristics, which are mental shortcuts used for decision-making in complex social environments, and highlights their advantages and disadvantages. It examines how these heuristics function, their origins linked to natural and sexual selection, and their implications for technology and autonomous agents. Various examples of heuristics, such as anchoring and representativeness, illustrate how they can lead to both effective and flawed judgments in everyday situations.
Introduction to Human Heuristics
- 1. www.aware-project.eu Introduction to Human Heuristics Material for social and pervasive computing Franco Bagnoli & Andrea Guazzini Center for the Study of Complex Dynamics University of Firenze, Italy www.complexworld.net
- 2. Introduction Humans do not deal with problems in a “rational” way. They use “rules of thumb” called heuristics, which are more “economic” than full rationality, but sometimes fail spectacularly. Our brain has been selected in a social environment, and we have developed heuristics to solve social problems, in limited time, with limited computational capabilities and with limited information available. Autonomous agents and portable devices are often confronted with similar situations, so the adaptation of human decision systems to computer science might be fruitful. Moreover, autonomous devices have often to collaborate with humans, and even act in their delegation. Franco Bagnoli & Andrea Guazzini (CSDC) Introduction to Human Heuristics 2 / 22
- 3. Are humans smart? Humans love to think to be intelligent and to take rational decisions. Actually, rational thinking is quite slow and computational demanding. We can discriminate the “usage” of cognitive capabilities by fMRI and response times. For instance, a good ping-pong player never “thinks” to the next move. Some partially “blind” people (blind sight) can detect movements even if they cannot “understand” what they see. Human recognition need “emotional” components, otherwise the subjects cannot even recognise themselves in a mirror. The signals that initiate a voluntary movement starts about 0.35 s earlier than the subject’s reported conscious awareness that he/she is feeling the desire to make a movement. Do we have free will in the initiation of our movements? Since subjects were able to prevent intended movement at the last moment, we surely do have a veto possibility. Franco Bagnoli & Andrea Guazzini (CSDC) Introduction to Human Heuristics 3 / 22
- 4. Heuristics as weak intelligence We have to take a lot of decisions in everyday life. Generally, these decision are satisfactory, but we all experience frustration for having chosen the bad choice, or having been cheated. Twerski and Kahneman examined many situations, and pointed out the existence of heuristics: “rules of thumb” that are used everyday, like for instance “prejudicial judgements” based on appearances. Clearly, if applied to a wrong context, heuristics may fail spectacularly. Heuristics may be hard-coded (and therefore sometimes called schemes) or learned. Franco Bagnoli & Andrea Guazzini (CSDC) Introduction to Human Heuristics 4 / 22
- 5. Examples of classic heuristics: anchoring When taking a decision, we rarely “weight” all factors, and generally rely heavily on just one piece of information (the one easier to recall), and only in a second moment we “adjust” the answer according to other factors. A classical example is the question “Estimate the probability of death by lung cancer and by vehicle accidents”. People tends to assign a higher probability to car accidents (since they are much more commonly reported by press) but lung cancer causes about 3 times more deaths than cars. If one asks if Turkey population is more or less than 30 million, and then asks to estimate that population, the average will be around that figure (Turkey has about 75 million population). Franco Bagnoli & Andrea Guazzini (CSDC) Introduction to Human Heuristics 5 / 22
- 6. Representativeness People are insensitive to prior probability of outcomes They ignore preexisting distribution of categories or base rate frequencies. Bayes’ theorem is not easily understood. People are insensitive to sample size They draw strong inferences from small number of cases People have a misconception of chance: gambler’s fallacy. They think chance will “correct” a series of “rare” events. People have a misconception of regression. they deny chance as a factor causing extreme outcome. Franco Bagnoli & Andrea Guazzini (CSDC) Introduction to Human Heuristics 6 / 22
- 7. Representativeness examples Is the roulette sequence “6, 6, 6” more or less probable than “10, 27, 36”? All kind of stereotypes: black people vs. white people, immigrants, etc. There is a murder in New York, and the DNA test (say 99.99% accuracy both for false positive and false negatives) is positive for the defendant. There are no other cues. Which is the probability that the defendant is guilty? Franco Bagnoli & Andrea Guazzini (CSDC) Introduction to Human Heuristics 7 / 22
- 8. Heuristics as fast and frugal processing At present, heuristics have a better reputation: they can be considered as “optimized” methods of saving computational resourced and giving faster answers (Gigerenzer). Many everyday problems would require “unbounded” rationality to be solved, and a large time for samplig all possibilities. But we do not try every possible partner when choosing a mate (nor a tiny fraction of them...). In a variable world, sometimes the “rules of thumb” are really better then the weighted methods taught by economists. In real world, with redundant information, Bayes’ theorem and “rational” algorithms quickly become mathematically complex and computationally intractable. Franco Bagnoli & Andrea Guazzini (CSDC) Introduction to Human Heuristics 8 / 22
- 9. A new view of heuristics Ecologically rational (that is, they exploit structures of information in the environment). Founded in evolved psychological capacities such as memory and the perceptual system. Simple enough to operate effectively when time, knowledge, and computational might are limited. Precise enough to be modelled computationally Powerful enough to model both good and poor reasoning. (Goldstein & Gigerenzer, 2004) Franco Bagnoli & Andrea Guazzini (CSDC) Introduction to Human Heuristics 9 / 22
- 10. Recognition heuristics In 1991 Gigerenzer and Goldstein asked twelve students in California and Germany to estimate whether S. Diego or S. Antonio had a larger population. German students were much more accurate, simply because most of them did not know S. Antonio. The same test was performed on soccer outcome, financial estimates, etc. But Oppenheim (2003) showed that we use also other cues. If asked to judge between a known little city and a fictitious one, most of people would choose the non-existing city. In any case, there is information in ignorance (and probably advantages in forgetting). Franco Bagnoli & Andrea Guazzini (CSDC) Introduction to Human Heuristics 10 / 22
- 11. Take the best We often have to choose the “best” (buy a new car). The most rational thing to do is to maximise a weighted score. The weights can be extracted by past experiences. For instance: you are a physicians and have to decide whether a man with severe chest pain should be sent to the coronary care unit or a regular nursing bed. The method based on weighted decision was slow, and had an efficiency of nearly 50% (i.e., random choice). A simpler decision tree is much more effective: first consider the most important factor – had the patient already experienced hart attacks? If yes, go to intensive unit. Then the second: is the pain localized in chest? If yes, go to intensive unit, etc. etc. This is why advertisers focus on “irrelevant” details for selling cars... Franco Bagnoli & Andrea Guazzini (CSDC) Introduction to Human Heuristics 11 / 22
- 12. Where do heuristics come from? Heuristics, like all our brain, is a product of selection. We are at hand with natural selection, i.e., competition for surviving. But in order to select a trait in this way, nature has to literally kill everyone not carrying that trait before reproductive age. A much less cruel but more effective selection is the sexual one. In many species, just a tiny fraction of individuals (the leading male, for instance) do actually reproduce. Franco Bagnoli & Andrea Guazzini (CSDC) Introduction to Human Heuristics 12 / 22
- 13. Sexual selection Sexual selection is so effective, that a tiny improvement in attracting the opposite sex can result in larger offspring. This is the origin of the extreme sexual ornaments found in all sexually-reproducing species. For humans, the principal ornaments are (probably) power and dexterity (mainly linguistic): poetry, songs,... It has been suggested that our “large” brain (with art and all useless brain products) is just a sexual ornament. Franco Bagnoli & Andrea Guazzini (CSDC) Introduction to Human Heuristics 13 / 22
- 14. Machiavellic brain Monkey and ape societies are often complex social systems. In such cases, the leading position is conquered by means of alliances, not by pure muscle power. This implies large cognitive power, since one needs to elaborate not only information about others, but also their mutual relationships. Actually, the size of frontal cortex (the “monkey” brain) correlates well with the group size (from which one obtains the Dunbar number for the human group size). Franco Bagnoli & Andrea Guazzini (CSDC) Introduction to Human Heuristics 14 / 22
- 15. Logic brain We find logic problems hard. How many cards should one turn (at minimum) to check if the following rule is violated? Cards with odd digits have a vocal on the back. Franco Bagnoli & Andrea Guazzini (CSDC) Introduction to Human Heuristics 15 / 22
- 16. Social brain But social tasks are easier... How many cards should one turn (at minimum) to check if the following rule is violated? People less than 18 cannot drink alcohol. Franco Bagnoli & Andrea Guazzini (CSDC) Introduction to Human Heuristics 16 / 22
- 17. Cooperative brain We have developed sophisticated methods for eliciting cooperation and punishing defeaters. Not surprisingly, this opens the way to (repeated) game theory... Franco Bagnoli & Andrea Guazzini (CSDC) Introduction to Human Heuristics 17 / 22
- 18. Example: the ultimatum game In this game, you are given 10$, and you have to decide how many dollars you will offer to a third person. He/she can accept and you share the money, or he/she can refuse and in this case both of you will loose everything. How much would you offer? If you were the third person, up to how much would you accept? What is the most rational thing to do? Franco Bagnoli & Andrea Guazzini (CSDC) Introduction to Human Heuristics 18 / 22
- 19. The dictator game This is the same as the ultimatum, but in this case the third person cannot refuse. How much would you offer in this case? Before answering, consider the following possibilities: This third person is sitting near to you. This third person is somewhere far from you. You personally know this person and you know that in some future time he/she can play you present role. You know that you’ll never meet again this person. You know that your choice will be made public in your school/office. What is the most rational thing to do? Franco Bagnoli & Andrea Guazzini (CSDC) Introduction to Human Heuristics 19 / 22
- 20. The trust game This is the same as the dictator, but in this case the third person initially offers some amount of money, which is doubled by the game manager. The dictator can decide to give back (partially) or keep for him/her-self. How much would you offer initially in this case (third person initial move)? Suppose you are offered 5$, which become 10. How much would offer back if you were the dictator? What is the most rational thing to do? Franco Bagnoli & Andrea Guazzini (CSDC) Introduction to Human Heuristics 20 / 22
- 21. This is the end... There is a network of nodes that process information coming from neighbors. The information can be corrupted, and in this case also the elaborated information is tainted, like a disease. The node remains infected only for a limited time. A node can check the correctness of the received information on a central repository, but it is costly (say, it takes time). Try to develop an heuristic for deciding when information should be checked. What additional information might be useful for reducing the infection level while not wasting resources in consultations? The “trustability” of neighbours? The average level of infection? How long should the memory last? How do these solutions depend on the geometry of the network? What does happen on a regular lattice/disordered graph/scale free networks? Franco Bagnoli & Andrea Guazzini (CSDC) Introduction to Human Heuristics 21 / 22
- 22. E-mail your proposal to franco.bagnoli@complexworld.net