Decision-Making and the Cognitive Architecture of Problem Solving

Decision-Making and the Cognitive Architecture of Problem Solving B. Chandrasekaran Keynote Talk First IEEE Symposium on Computational Intelligence in Multi-Criteria Decision-Making (MCDM 2007), Honolulu, HI April 3, 2007 Research supported through participation in the Advanced Decision Architectures Collaborative Technology Alliance sponsored by the U. S. Army Research Laboratory under Cooperative Agreement DAAD 19 -01 -2 -0009. Copyright © 2007 by B. Chandrasekaran

Background for the Talk • My long interest in cognition, cognitive architectures • Relatively recent interest & work in Multi. Criterial Decision Making (MCDM) • Puzzles about the relation between the general human situation in acting and deciding and the highly formal normative formulations in decision theory.

Decision- Making in the Wild (DMITW) • DMITW is everyday DM, before the formalization that potentially makes the problem capable of a “rational” solution, the formal version. – Often formulated as maximizing utility (u), or satisfaction of the agent. • Ex: A man wins $20, 000 in a lottery. Should he pay off some of his loans? Buy a new car? Repair house? Maybe other things? The formal core may be MCDM, but: – Choices (alternatives -- “alts”), criteria not given. DM has to figure them out. – In any case, this is one problem among many faced by agent.

Pre-Formal version process itself subject to informal rationality judgments • Rationality in DMITW is also to be judged by the degree to which the process results maximizing utility for the agent, given agent’s knowledge. • Many ways of falling short: – Less complete set of alt’s • “I could’ve had a V 8. ” Failed to generate the V 8 alt. Why would if have been more rational to have thought of V 8? • How to increase the likelihood that the V 8 alt is generated? – Poor evaluation of alt’s • How to ensure that likely satisfaction is estimated correctly? – Choosing the “best” alt • Getting this right is the easy part. Once formalized, normative solutions are available. .

My Focus • I wish to focus on phenomena in this region, esp. the role of cognitive architecture in transforming the problem into the “formal” version. – The strengths & weaknesses of cognitive architecture then become central. – Analysis may help understand various “irrationalities” and how DSS’s may be enhanced.

Bias & “Irrationality” • Human decision-making is subject to various irrationalities. Examples: – Framing Bias – Transitivity issues, – “Fairness” bias, … • Extensive literature, I’ll just touch on a few aspects. • Understanding their origin and help in designing DSS’s that can help DM’s avoid them.

Bias & “Irrationality” Examples: Framing Bias • Subjects received a message indicating that they'd initially receive £ 50. • They had to choose between a "sure" option and a "gamble" option. The "sure" option was presented in two different ways, or "frames" to different groups of subjects. • The "sure" option was formulated as either "keep £ 20" (Gain frame) or "lose £ 30" (Loss frame). The "gamble" option was identical in both frames and presented as a pie chart depicting the probability of winning.

Bias & “Irrationality” Examples: Framing Bias

Morgenbesser’ Pies • Anecdote about Philosopher Sidney Morgenbesser: – After finishing dinner, Sidney decides to order dessert. – The waitress tells him he has two choices: apple pie and blueberry pie. He orders the apple pie. – After a few minutes the waitress returns and says that they also have cherry pie at which point Morgenbesser says – “In that case I’ll have the blueberry pie. ” • This was supposedly a joke, but might the above behavior actually be rational?

Cognitive Architecture & Problem Solving • Cognition can be analyzed architecturally in many layers. • We focus on the problem solving (PS) level of analysis, such as embodied in Soar: – Goals, result in relevant knowledge being accessed from LTM, which sets up a problem space, which may result in subgoals, and this is a recursive process, until all the subgoals are achieved, & the main goal is as well. – Open-ended, search in problem spaces. – Heuristics & pre-compiled solutions can reduce search – Aspects of the architecture may bias the search in various ways.

Evaluating Alternatives by Problem Solving: $20, 000 Windfall Example • Knowledge used to generate alt’s (not shown), evaluate consequences

Reducing Search • Directly relevant knowledge – e. g. , result of learning ftom previous situations, can reduce search. • Heuristics. Similar to above, but less dependable – The heuristics may directly assign “desirability” values based on elements of alt descriptions. • Architectural biases: – Avoid loss, Avoid danger, built-in notions of fairness, • Set by evolution – These result in rapid evaluations of alts.

Ideally, PS transforms a DMITW into a formal version with a normative solution. • Given a decision problem, the DM : – Simulates, imagines, calculates, processes, composes – I. e. , Problem Solves - - until she can get a sense of the possible alternatives, and their consequences. • One approach is to transform the problem into MCDM. – Proceed with simulating consequences until each alt is characterized in terms of similar dimensions, I. e. , “criteria. ” • Not always possible, but when possible, the alt’s can be compared more readily. – But how does MCDM connect to utility in general?

Example: $20, 000 windfall • Knowledge used to generate alt’s – Pay off debt (POD) vs fix house (FH) – POD will reduce monthly interest payments, increase net income; FH … Inc. in income POD FH Domestic peace Conscience improvement

MCDM and Utility • In MCDM, the DM is normally characterized as one with multiple objectives. – But we need a unification with the more general “DM with utility” model. • Let Ci(dk) be the value alt dk has on the criterion Ci (w/o loss of generality, all criteria directions are such that higher is better), and u(dk) the utility of alt dk. • The dimensions -- the criteria – should be as informative about u as possible: – f(C 1, C 2, …Cn) uest

Criteria & the u function, contd • i. Ci(dk) > Ci(dl) --> Pr[u(dk) > u(dl)| Ci’s] (call it pi) is greater than ½. Alt’s can be ordered with some confidence. • ii. Ci(dk) > Ci(dl) & Cj(dk) > Cj(dl) --> Pr[u(dk) > u(dl)| Ci’s & Cj’s ] (call it pij) is greater than or equal to pi. Alt’s can be ordered with more confidence. • iii. If Ci(dk) > Ci(dl) & Cj(dk) < Cj(dl), the alt’s cannot be ordered, and more knowledge is needed. – A trade-off situation. Perhaps additional dimensions of comparison can be identified.

Trade-off’s • Trade-off’s are normally described as the agent “accessing” the preferences. In fact, in DMITW, they typically require additional simulation, imagining, etc. – More descriptive dimensions, giving more into about relative values of u’s.

Possible Sources of Error: Cognitive Architecture & Bounded Rationality • Knowledge limitation: generation of alts, exploration of consequences may both suffer. – What knowledge is invoked depends on the terms in the problem description and on how the alts are described. Different descriptions, though logically identical, may invoke different knowledge. One source of Framing Bias. • Architectural limits & biases: – STM limitations – Hard-wired architectural features might bias evaluations & choices • “loss, ” “danger, ” in alt descriptions may influence assessments of alts. Another source of Framing Bias. • Time limitation: Incomplete exploration, more dependence on heuristics, architectural biases.

Cognitive Architecture & Explanations of Biases • Some of the “irrationalities” may be explained by the features of the cognitive architecture. • Not everything that appears irrational may be so if viewed in the larger context.

Bias & “Irrationality” Examples: Framing Bias • Subjects received a message indicating that they'd initially receive £ 50. They had to choose between a "sure" option and a "gamble" option. The "sure" option was presented in two different ways, or "frames" to different groups of subjects. The "sure" option was formulated as either "keep £ 20" (Gain frame) or "lose £ 30" (Loss frame). The "gamble" option was identical in both frames and presented as a pie chart depicting the probability of winning.

Architectural Bias against “Loss” uest(Lose $30) < uest(Keep $20) • Evaluating “Bet” requires a bit of problem solving • Issue (probably) does not arise if choices are: – {Lose $30, Lose $20} vs {Keep $20, Keep $30} – But how about: {Lose $30 Vs Keep $30}?

Possible Rationality of Prof. Morgenbesser’s Behavior • Recall: – The waitress gives two choices: apple pie and blueberry pie. He orders the apple pie. – The waitress returns and says that they also have cherry pie at which point Morgenbesser says: “In that case I’ll have the blueberry pie. ” • Can a problem solving model explain this? – How about if the “cherry pie” alt triggered new knowledge that generated new dimensions of comparison that had not previously been taken into account?

Framing Bias Related to Knowledge Invocation • Consider example in sensor-allocation planning, where some of the criteria are: “coverage, ” “timeliness, ” and “fuel expended. ” The last criterion can also be described by “fuel remaining” – one may be computed from the other. – When doing trade-offs between two COA’s, The DM may have to simulate alt’s further to resolve trade-off. – The different descriptions can invoke different problem spaces: fuel remaining may bring up problem spaces about future missions while fuel expended brings spaces about cost for current mission. The two descriptions may result in different trade-off behavior. • Depending on the larger context, both descriptions may be used for better effect, though it may appear redundant.

Bias & “Irrationality” Examples: Intransitivity • Arrow’s Paradox in Voting (collective decision-making) – Versions exist in individual DM as well. – Less a result of hard-wired properties of the architecture and more one that arises due to discarding relevant information. • Counting positives & negatives. • Possible consequence of knowledge limitation, discretization of evaluations.

Bias & “Irrationality” Examples: Fairness Bias • Players A & B. Both A & B know the following rules. • A is given $100 and asked to share with B. A can propose to B any amount to share. If B accepts, he gets that amount, and A gets to keep the balance. If B declines, neither of them get anything (A has to return the $100). • By a significant margin, people asked to be B reject “unfair” offers, even though they’ll end up getting nothing. • This bias is likely architectural.

Architectural Biases in Moral Decisions • • Peter Singer, March 20, 2007, The Guardian (UK) Greene. . . studied how people respond to a set of imaginary dilemmas. In one, you are standing by a railroad track when you notice that a trolley, with no one aboard, is heading for a group of five people. They will all be killed if it continues on its current track. The only thing you can do to prevent these five deaths is to throw a switch that will divert the trolley on to a side track, where it will kill only one person. . . most people say you should divert the trolley on to the side track, thus saving a net four lives. . . In another dilemma, the trolley is about to kill five people. This time, you are standing on a footbridge above the track. You cannot divert the trolley. You consider jumping off the bridge, in front of the trolley, thus sacrificing yourself to save the people in danger, but you realise you are too light to stop the trolley. Standing next to you is a very large stranger. The only way you can prevent the trolley from killing five people is by pushing this stranger off the bridge into the path of the trolley. He will be killed, but you will save the other five. --. . . , most people say that it would be wrong to push the stranger. . . . remarkable consistency despite differences in nationality, ethnicity, religion, age and sex. . Greene found that people asked to make a moral judgment about "personal" violations, like pushing the stranger off the footbridge, showed increased activity in areas of the brain associated with emotions. This was not the case with people asked to make judgments about relatively "impersonal" violations like throwing a switch. Moreover, the minority of subjects who did consider that it would be right to push the stranger off the footbridge took longer to reach this judgment than those who said that doing so would be wrong. .

What Does All this Mean for DSS’s? • A decision is only as good as the quality of the formal version of the problem, however “rational” the solution is wrt the latter. • Ultimately, all that the human DM has is hir cognitive architecture, which has its strengths & weaknesses. – Strengths: • Large, potentially open-ended amounts of knowledge, including knowledge about accessing external knowledge. • Ability to bring relevant knowledge, if very slowly, to bear on the task at hand. – Weaknesses: • Slow. • Evolutionarily hard-wired architectural constraints & biases. (framing a fairness biases) • Access to knowledge depends on cues in descriptions & may miss relevant knowledge (framing bias).

What Does All this Mean for DSS’s? • DSS design should be sensitive to possible DM biases arising from the architecture. • DSS’s can help here. – Framing Bias: Investigate the operational context to make sure descriptions help the DM evoke all relevant knowledge.

How DS’s can Help • Supporting exploration of problem representation & formulation • Supporting alt generation techniques humans are poor or slow at: – E. g. , GA techniques for generation of design candidates, candidate composition techniques – … • Supporting evaluation techniques humans are slow or poor at: – E. g. , simulation

Questions?