Experience and Knowledge

Why is it that with every passing year, I feel experience is more important than ever for solving problems—even after half a century? Why isn’t knowledge enough? By knowledge here, I mean the kind that is typically called “book knowledge”; the set of information, facts, How-To’s that one might read or be taught in a classroom. The Internet and search engines are probably the greatest example of such repository of book knowledge. With the content of the Internet at my fingertips, I should be able to make as good decision as the most seasoned veteran of…well, whatever it may be.

Of course, some kinds of knowledge simply cannot be transferred without direct experience, such as action under extreme stress and threat, personal emotional states like flush of attraction, etc. Other kinds of knowledge require practiced expertise such as hitting a baseball. What is curious is that in many situations where emotional response or physical learning should play a lesser role and where transfer of information should seem straightforward, experience seems to still matter greatly.

It is pretty easy to look back and recall various errors we might have avoided if, with the same knowledge, we had more experience. This state of affairs is obviously true for teen-age years, quite true for your college years, and usually true for late 20’s and early 30’s. What is interesting to me is that after 50 years of experience, I still feel that there are problems, jobs, situations, where I could have avoided certain errors or done things differently had I more experience—again, assuming the same set of book knowledge. What is it that experience represents in problem solving?

Many tasks and jobs in the world require hands-on training and experience. I currently “run a lab” to carryout biomedical research. Much of the operations of a lab are written in manuals and papers, but much of it is not. Even for experimental protocols that are written, say for sequencing DNA, many details are missing. Even if you read all the laboratory protocols for DNA sequencing experiments and have that knowledge at your finger tips, it is doubtful that you would be able to successfully carryout the experiment if I left you alone in a lab. In this case, a large part of the problem is that the written information is quite incomplete. Laboratory manuals are notoriously sparse, assuming a great deal of prior knowledge. Primary papers in journals are much worse. Repeating some experiment purely from a paper is nearly impossible. I suspect this is the general case in all fields of high expertise. Whenever I see a new building construction site, I marvel at the complexity that must be involved and wonder how much of the knowledge is written and what is handed down by demonstration.

It may be that with a medium like videos, information can be conveyed more efficiently than with a medium like books. This is probably the great hope of the Internet open course movement. A picture paints a thousand words and all that. Of course, such a video needs to be accompanied by a lot of explicit information. Suppose in handling a microtube a scientist in the video twitched a few times because of an itch—a student learning from the video might think shaking the tube is part of the protocol. This is not so stupid. There are many examples of experimental protocols where some step comes from ritualization of an initial accident.

Incomplete information might be ameliorated by recording and transmitting more information. Manuals, instructions, lectures can be made much more detailed. This would be a great help in many cases. On the other hand, the same sentence may be understood differently by different people and there is really no way to write something for universal comprehension--which is why we need actual classrooms where real-time feedback helps to individualize the communication. But, that is a topic for another day.

Suppose we take all of the caveats above into account so that we have all the knowledge at our finger tips, the problem doesn’t require practice, and it doesn’t involve emotional participation—does experience still matter? As I mentioned, after 50 years of living and 25 of them in a professional career, I still feel strongly that each experience gives me insight that I did not realize prior to the event. Suppose we narrow it down to the case where we are presented with a “problem situation” and we have to chart a series of actions to solve the problem. In the first step, we need to recognize the “situation” and classify it such that it maps to the possible action items. For example, given a laboratory and the problem of sequencing some DNA, I have to map the problem to the input of each item in the lab (e.g., the black square box is a PCR machine that loads small plastic tubes). While we can have tremendous amount of knowledge, we cannot have infinite amount of knowledge. Thus, the problem of recognition and mapping to actions typically requires some kind of extrapolation/interpolation. That is, I need to recognize that a thing with a metal rack that heats up and a LCD screen interface is likely to be a PCR machine that accepts tubes of liquid, even if I didn’t have a record of all possible PCR machines. Second, once I recognize the problem and how it might be mapped to the input of various actions, I have to chart the sequence(s) of actions. We may have very many possible variations and very many parameters for each action. Charting an optimal solution will require considering an extremely large number of combinatorial possibilities—maybe infinite. Lastly, we have to decide on an appropriate optimality consideration. That is, we want the solution to be optimal as measured by an (so-called) objective function. The possible objective functions might include valuation of accuracy, time, impact, etc., or some combination of these. It is common that different objective functions cannot be simultaneously optimized.

Leonid Kantorovich, the only Soviet Nobel Laureate in economics, developed the theory of optimal allocation, especially with respect to a managed economy like that of the former USSR. Any reasonable sized economy has millions of input and outputs and associated parameters. Is it possible to optimize over these variables to maximize productivity? It doesn’t take too much to see that even if we were able to formally model the problem, the optimization problem could not be realistically solved. Markets are heuristic local greedy optimizers to the resource allocation problem. It is what we use not because it is the optimal algorithm but because it is a practical heuristic. A heuristic algorithm is somewhat hard to define. It tries to solve an optimization problem but it cannot guarantee the solution or even a bound on the solution. By a bound, I mean statements like “the solution is no worse than 2X the optimal solution.” Heuristic algorithms typically contain rules that seem sensible vis-à-vis the problem, e.g., make finite number of players bid for a resource and allocate the resource to the highest bidder, but the overall strategy cannot be theoretically analyzed to give guarantees.

The utility of a heuristic algorithm cannot be theoretically proven but it is usually validated with empirical testing—that is, by experience. Human chess players do not use the same kind of algorithm as computer chess programs. But, many players can beat most computer programs most of the time, demonstrating the utility of human heuristics. Most of these heuristics are self-taught by experience. So then, for problems in general, just like the specialized case of chess, we learn heuristic strategies for finding the optimal solution. Only heuristics is possible because the complexity of the problems do not admit a global solution within the allocated finite time. More experience may give us better heuristics. Finally, as mentioned, there are many possible objective functions—the choice of the right objective function may be governed by some meta-objective function, say “happiness”. The heuristics for optimizing such meta-objective function may be involved in what we call “wisdom.”