Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Intuition and Deliberation - Two Systems for Strategizing in the Brain Duo-process theory distinguishes between intuition and reasoning. Intuition (system 1): fast and emotional Reasoning (system 2): slow and controlled These differences imply different games might be fundamentally different. coordination games vs. dominance solvable games dominance solvable games prisoner’s dilemma game eliminate Confess Don’t confess Confess 2,2 4,0 Don’t confess 0,4 3,3 iterated elimination of dominated strategies eliminate (first) eliminate (second) Confess Don’t confess Confess 2,2 1,0 Don’t confess 0,4 3,3 Games that can be solved by iterated elimination is called dominance solvable games. Notice that computers can be very good at dominance solvable games, maybe even better than human. Coordination games are very different. Meeting in NY game Grand Central Empire State Grand Central 1,1 0,0 Empire State 0,0 1,1 There is nothing to eliminate. Computers may be very bad at it but human may be better. Studies on coordination games are few. Mehta et al (AER, 1994) asked subjects to name one color, one number and one year. •When subjects were not rewarded based on their answers, they… •When subjects were rewarded if they said the same thing as their randomly assigned partners, they… •When subjects were not rewarded based on their answers, blue and red turned out to be equally popular (35% each), the most popular number was 7 (11%) and the current year was named 6.8% of times. •When subjects were rewarded if they said the same thing as their randomly assigned partners, 58.9% people named red, 40% of people chose number 1 and the 61.1% named the current year. In Schelling’s terms, the color red, the number 1 and the current year are focal points. •What exactly focal points are and how do people reach a focal point partly motivate this study. •We would like to study both the behavioral and neural difference between dominance solvable and coordination games. Compare coordination games with dominance solvable games: There is nothing to eliminate. Computers may be very bad at it but human may be better. Methodology: We scanned 21 subjects in an fMRI scanner when they were making choices in dominance solvable and coordination games. •We would like to know the natural reaction of subjects so only subjects who have never taken any game theory course were recruited. •Moreover, no feedback was given inside the scanner. •21 is a good sample size. Most importantly, the payoff matrix is hard to grasp (we cannot do it without paper and pen and it is not a good reason if equilibrium fails just because people have difficulties in understanding the payoff matrix rather than in figuring out an equilibrium), so we use the following new design of box games and number games. 2 treatments: box and number 2 conditions: domi and coor The following are some sample screens that subjects saw. 1 2 3 A B C you:1R other:1R 1U 1 2 3 A B C you:other other:you You: 0 1 2 3 Other: 0 1 2 3 you:other + 1 other:you You: 0 1 2 3 Other: 0 1 2 3 you:other other:you Let us now solve for the two sample dominance solvable games. Thank god there is nothing to solve for coordination games. you:one right other:one right one up 0 1 2 3 0 1 2 3 0 1 2 3 0 1 2 3 0 1 2 3 0 1 2 3 0 1 2 3 0 1 2 3 you:other + 1 other:you 0 1 2 3 0 1 2 3 0 1 2 3 0 1 2 3 0 1 2 3 Some of our subjects do like this (go back and forth between herself and the other). The point is it is easier to do elimination by eyes for these games. ● ● ● ● ● you:one right other:one right one up ● 0 1 2 3 0 1 2 3 ● 0 1 2 3 ● 0 1 2 3 you:other + 1 other:you ● 0 1 2 3 ● 0 1 2 3 Now let us go back to the sample coordination games. See how the answers “pop out.” Our fMRI subjects (21) respond by: Our pilot subjects (41) respond by: Domi Coor 1 (1) 1 (3) 16(e) (32) 2 (4) (1) 3 (1) 18 (33) (2) (4) Domi Coor 0 1 2 3 1 1 1 18 (1 5 2 33 0 1 2 3 18 2 1 0 (e) 12 8 14 4) According to Schelling, focal points “have symbolic or connotative characteristics that transcend the mathematical structure of the game.” We have some interesting sample questions in this respect. You: 747 767 380 757 Other: 747 767 380 757 You: 911 228 921 124 Other: 911 228 921 124 1 2 3 4 A B C D experiment design and time line: event related and self-paced Coor or Domi Coor or Domi Coor or Domi Now we know the games, we turn to the fMRI (functional magnetic resonance imaging). a scanner a typical EPI image When neurons fire, blood releases oxygen to them at a greater rate than to inactive neurons. Since oxyhemoglobin and deoxyhemoglobin have different magnetic susceptibility, this magnetic signal variation can be detected using an MRI scanner. Given many repetitions of a thought, action or experience, statistical methods can be used to determine the areas of the brains which have more of this difference as a result, and therefore which areas of brains are active during that thought, action or experience. MR signal (S) active resting t TE excitation reception experiment parameters Slice 25, FOV read 256, FOV Phase 100 Slice Thickness: 4mm TR: 2000 ms TE: 34 ms Phase Oversampling: 38 Base Resolution: 64 Phase Resolution: 100 In a nutshell, we estimate a general linear model. Signal = bd[Domi]+bc[Coor]+bm[Motor]+ε and we run t-tests on bd - bc and bc- bd. First, let us look at our behavioral data. expected payoff avg. payoff against mode Domi 35.81 sec (15.98) 66.97% (21%) 79.26% (21%) Coor 9.78 sec (2.90) 46.10% (18%) 69.92% (18%) reaction time 1.0 expected payoffs Average of participants' 0.8 0.6 0.4 0.2 0.0 Box-Game Number-Game Grand Total Dominance Solvable Coordination 70 Average response times (secs) 60 50 40 30 20 10 0 Box-Game Number-Game Grand Total Dominance Solvable Coordination Two observations can be drawn from the behavioral data. •Humans are really different from computers (esp in Coor games). If we test whether subjects’ choices are random by Pearson test, the p-value is 0.00 (0.00 for Coor). So they are not choosing randomly. Bootstrapping the distribution of expected payoffs shows that the p-value of the expected payoff of a randomizer is 0.00. •They take longer time to make choices in Domi games than in Coor games. The difference in reaction time complicates data analysis. One can argue that the difference we see in bd – bc and bc- bd results from the difference in reaction time instead of the difference in nature of these two types of games. To address this, we run three other regressions. In one, the “task” regressor is parametrically modulated by reaction time. In another, reaction time is incorporated as other regressors (which do not convolve with HRF). In the third, we only model the first 8 seconds of each trial. These three regressions give very similar results. Let first review some basics about the brain. Domi-Coor (bd- bc) Coor-Domi (bc- bd) In short, in Domi-Coor, we find activation in fronto-parietal network (the middle frontal gyrus, the inferior parietal lobule, precuneus). On the other hand, in Coor-Domi we consistently find activation in insula and ACC. Domi-Coor Coor-Domi What exactly do these results mean? Domi-Coor: working memory Coor-Domi: salience detection Domi-Coor Fronto-parietal activation results when tasks require attention, conscious perception, reasoning and memorizing. Domi-Coor Previous literature: fronto-parietal activity observed when contrasting logical reasoning to tasks where reasoning is not required, when contrasting challenging reasoning tasks to straightforward ones, or when contrasting a meaningful middle game position to a random game position. Dominance solvable games may induce players to go through steps of reasoning. In each step, players may need to eliminate some choices and memorize. Domi-Coor Verbally encode and hold the targets of both: Inferior parietal lobule is implicated in verbal memory storage. Eliminating dominated strategies may engage the central executive which is to manipulate the contents of storage: Middle frontal gyrus is thought to execute goal-directed operations. Keeping track of which strategies are eliminated may require generating and retrieving a mental image: Precuneus may be related to memory retrieval and imagery. Domi-Coor The greater fronto-parietal activation in dominance solvable games is consistent with the hypothesis that these areas assist in step-by-step deliberative mental processes. One anecdote is this. The parietal lobe of Einstein’s brain was 15 percent larger than average. Maybe that is why he is smart? Coor-Domi Insula: implicated in subjective pre-reflective and reflective representations of ongoing changes in internal bodily and feeling states. Insula receives information from receptors in the skin and in internal organs. If an animal is hot, it seeks shade. If hungry, it looks for food. If hurt, it licks the wound. ACC: a conflict monitor when tasks require attention, novel or open-ended responses or when cognitive uncertainty exists. Coor-Domi Previous literature: When participants contemplate cooperating instead of competing with another person, when they judge other persons to be trustworthy instead of being untrustworthy, or when they experience social emotions such as empathy or love, insula and ACC activate. These emotions might have evolved to ensure quick responses to the factors arousing them in the presence of many stimuli. Coor-Domi Many social interactions involve myriad stimuli but demand immediate decisions. Rapid processing and extraction of the most salient aspects of complex situations is characteristic of intuitive decision making. Complex: multi-purpose, respond to many factors and be flexible Quick processing: extract the most salient feature Coor-Domi The insula and ACC seem to be part of a general network contributing to a quick and exible evaluation of complex multi dimensional experiences. In our experiment, deciding within the context of a coordination game which Nash equilibrium has the most salient characteristics requires rapid processing of various cultural connotations as well as geometric symmetry, centrality or even mathematical oddity. Coor-Domi The middle insula receives inputs regarding the physiological condition of the body from the posterior insula and integrates this with salient environmental stimuli. The posterior insula and the SMA/CMA are shown to be responsive to changes of many sensory modalities, whereas the anterior insula and ACC are sensitive to novelty. Coor-Domi The key to coordination may be the ability to make judgments of salience common to both. That is, the participant must identify features which are likely to be salient not only to herself but also to the other. The previous studies all point to a possible role of insula and ACC in identifying salience and provide support for the hypothesis that the higher activation we observe is due to participants extracting salient features in order to coordinate. Coor-Domi Post-scan interviews indicated that many used intuitions or “gut feelings” to identify the focal points of coordination games. They just know how to choose in coordination games. Maybe they get their gut feeling from Insula and ACC? Step To solve a dominance solvable game requires a well defined number of steps. The greater is the number of steps, the more taxing might be the mental processes. Maybe the fronto-parietal activation in dominance solvable games might correlate with the number of steps? NCI For coordination games, NCI (normalized coordination index) measures how well coordination is achieved. Choice 1 Choice 2 Choice1 Choice2 Choice 1 Choice 2 Group A 10 0 5 5 10 0 Group B 0 20 10 10 20 0 NCI (1*0+0*1)*2= 0 (0.5*0.5+0.5*0.5) *2=1 (1*1+0*0)*2= 2 0 1 2 NCI A high NCI means a high coordination rate, perhaps reflecting a strong gut feeling aroused by an obvious focal point while a low NCI (close to 1) means choice is close to random. Maybe the activation in insula and ACC in coordination games may correlate with the NCI? Step and NCI We divide 40 dominance solvable games into 20 hard and 20 easy games depending on the number of steps and similarly divide 40 coordination games into 20 highly focal games and 20 less focal games. We build a second model where the two kinds of games are modulated by the categories respectively or Signal = bd[Domi] bdStep[Domi Hard or Easy]+bc[Coor]+bcFocal[Coor Focal or Less Focal] +bm[Motor]+ε. Step and NCI Precuneus activates more in hard domi than in easy domi. Precuneus plays a role in memory retrieval. The higher demand for memory imagery and retrieval in hard dominance solvable games may explain this activation. Insula activates more in highly focal coor than in less focal coor. The participants may have felt quite strongly that the other must notice the same salient feature. This may be why insula activation is higher for games which are more focal. Payoff A difficult game may generate a lower reward and different participants may disagree about which games are difficult. We build still another model where each task is modulated by its grades, Signal = bd[Domi] +bdgrade[DomigGrade] bc[Coor]+bcgrade[CoorGrade] +bm[Motor]+ε. Payoff Precuneus activation is negatively correlated with domi grades of a subject. Domi games with lower payoffs present harder mental challenges. Insula activation is positively correlated with coor grades of a subject. Coor games with higher payoffs arouse the gut feeling stronger. Step, NCI, Payoff The second and the third model suggest that the more “effortful” mental process activated by dominance solvable games are more heavily taxed when the games are hard to solve, while the more “effortless” mental processes activated by coordination games are more strongly activated when coordination is easy. This provides additional support for the main hypothesis of our research, that is, dominance solvable games and coordination games set off two quite different mental processes. The big question is, if we know that there are two very different networks that are activated when people play Domi or Coor games, can we apply what we have learned here to predict behaviors in other games? For instance, in trust games, do people reason or rely on intuition? Can we infer from the brain activations to make an educated guess at what they will do? Can we categorize players when they play an interesting game (like p-beauty contest) by their brain data so that we know whether they are using a more calculating way or a more intuitive way to reach their choices? Hope you find the talk interesting.