While ideally aiming at understanding an empirical phenomenon, and at the same time predicting its behaviour, scientists often focus only on one of these two aspects. Is this the dilemma of a trade off, a psychological bias, or a philosophical conundrum?
It is as if A is asked a question by B, and tries to give the correct answer. A could aim at understanding what is in the mind of B, grasping the rationale of the question that leads to the answer. Alternatively, A could simply aim at predicting the correct answer, from other lines of reasoning (knowing B personally, remembering things B once said, imitating others that have already replied, replicating a pattern, etc.).
This underscores fundamental psychological and philosophical differences in the personal approach of scientists to knowledge. Indeed, it is possible to optimise either conceptual/formal or predictive aspects of scientific models. At a higher scale, these often correspond to the fundamental and applied methodological approaches, respectively.
Would we rather predict the behaviour of a phenomenon that we don't understand; or understand it, while being unable to accurately predict its behaviour?
Have you considered the possibility that the dichotomy you propose/perceive/allude to is perhaps false?
In my estimation, the best models serve at least 3 purposes - 1 to help abstract away from observations (data), provide (mechanistic) explanations and provide predictions that will lead to things hitherto unknown/unexplored, or tests that can help refine the model.
I am no philosopher of science, but to me the scientific value of a modelling exercise is contingent upon being able to satisfy all 3 purposes, perhaps to differing degrees. Maybe one comes up with a model that elegantly re-describes a lot of empirical observations and makes some non-trivial predictions, but loses explanatory power when queried at increasing levels of granularity. Maybe a model works to explain low-level nuts-and-bolts aspects of something highly complex but does not scale well, and its explanatory, descriptive and predictive powers are limited to a level of abstraction different from that of the phenomenon of interest.
I am intrigued by the last bit...could you give me an example of something "we understand" but cannot "predict"? I can think of many examples of the converse, but this is a bit odd. As for the "would we rather part'...I am all for understanding AND predicting as best as I can. If pushed I would be inclined to have more understanding despite less predictive power, but I am not sure I have thought of non-trivial cases where this would apply.
Dear Anand,
this is exactly the type of feedback I was looking for. I had this thought when I read an online lecture by K. McGarigal (Univ. Massachusets Amherst) on landscape models (http://www.umass.edu/landeco/teaching/landscape_ecology/schedule/chapter10_modeling.pdf):
"(...) the goals of the study often dictate which stages of model-building will receive the most emphasis. If the goal is to integrate information and generate a deeper understanding of the system, then the early stages of model-building are crucial (conceptualizing, formalizing) and model analysis is critical. Model tests might focus on evaluating alternative conceptual models or formalisms. If the goal is accurate predictions, then verification and validation against data are of paramount importance, and a variety of equally plausible (or even implausible!) model formalisms might provide accurate predictions in at least some cases."
I understand that the dichotomy is just describing the extremes of a spectrum, but undoubtedly there is often a psychological bias towards one of these goals, in the mind of the modeler.
In my example of A and B, A may understand the intentions, motivation, and rationale of the question asked by B, and yet not being able to predict correctly the answer. For instance, A might understand B is thirsty, that *he likes orange juice, that *he is trying to lose weight (thus likely not willing to drink a coke), and yet might be unable to correctly reply to B's question: "what I would like to drink right now?". The model of B's behavior can attempt at maximizing one or the other aspect. Perhaps we can focus on what B is looking at while *he poses the question, or what *he experienced in the last hour, and this would give us a higher power of prediction, than we could have if we focused on the understanding of her mental state, and h** psychological background. You can see here a clear dichotomy between an attempt of understanding the phenomenon (the psychology of B) and of predicting its behavior (answer to B's question), in the latter case essentially treating her psychology as a black box.
Can you see my point now?
You can combine both! For example, I created a mechanistic model of seed dispersal to predict seed dispersion patterns (including parameter values) directly from the different characteristics of the natural history of plant dispersal and behaviour of their dispersal agents, which also incorporated phenomenological empirical mathematical models :)
Andrew, does this imply that your model conceptualizes and formalizes the actual mechanism of seed dispersal, or that it simply uses correlations between seed dispersion patterns and behavior of dispersal agents?
Yes, the model actually showed movement of variable numbers of individual birds through landscapes of fixed abundance of fruiting plants whose spatial arrangement could be varied. The birds deposited seeds based on algorithms determined by gut passage time of seeds, bird perching time, and attraction to trees based on distance and fruiting level. You can read the paper on my page or I can send a copy :)
- Badges
- Science topic
- Similar topics
- Engineering
- Modeling