This week I gave a seminar on Antifragility to my colleagues at the Social Science Research Unit. The concept of Antifragility was developed by Nassim Nicholas Taleb and popularised through his “Antifragile: Things That Gain from Disorder” book. The seminar however was inspired by a separate essay, published by Taleb on Edge.org: Understanding is a poor substitute for convexity (antifragility). This essay specifically addresses the antifragile qualities of scientific research, and the consequences for science funding. As a result, it was pretty easy to adapt it to Social Science research (where, by definition, the number of significant variables is known to be unknown) and to its application in policy making.
This post is designed to make the content of the seminar available: the PowerPoint presentation is here.
The slides are annotated (in “View\Notes Pages”) to provide the necessary explanations. For convenience, the slide annotations are also reproduced below.
This is necessarily a very short introduction, a much longer discussion could be made. I will probably post again with more personal views on the subject, but for the time being I think it is a good idea to limit this post to the contents of the seminar alone. Comments on the presentation and/or the general idea are of course welcome.
Preparing this little seminar was highly enjoyable, and, as far as I can tell, the reception was excellent (but of course, I am not in the position to judge).
Slide 1 (Title):
Harvesting uncertainty: can we maximise research outcomes by exploiting the antifragile properties of science?
By Sergio Graziosi – Social Science Research Unit, Institute of Education, London.
This is the PowerPoint presentation used in the SSRU seminar, marginally re-adapted for the web:
1.Where appropriate, notes are included, in order to explain the content of each slide.
2.Most animations have been removed.
The seminar was delivered on 6 November 2013.
The views expressed here are those of the author and do not necessarily represent the views of, and should not be attributed to, the Social Science Research Unit or the Institute of Education.
Slide 2 (this is a Seminar/Workshop) & Slide 3 (Antifragility? What is this about?):
Slide 4 (Nassim Nicholas Taleb):
The seminar was inspired by Taleb’s essay published on Edge: http://www.edge.org/conversation/understanding-is-a-poor-substitute-for-convexity-antifragility.
The original essay deals with the antifragile properties of research as a whole, and mostly from the point of view of research funders. However it can be adapted to the special case of Social Science research, both from the perspective of the researcher and/or of policy makers.
Slide 5 (Antifragility? Does it even exist?):
A few examples to convince the audience that Antifragility is all around us, with the chance to explain its most relevant properties:
1. The distinction Fragile/Robust/Antifragile always applies to a specific domain. For example “Transport in London” refers to reliability and relative speed (when compared with the other options), and how these change depending on what is happening in London. On the domain of personal safety, the bicycle would be fragile, but this doesn’t negate its ability to adapt to different traffic conditions and to improve it’s usefulness when some random event affects the other systems.
2. The categories on the right column benefit from volatility, chance and the unexpected. For example, if the reputation of a commentator (journalist or similar) is highly controversial, s/he will gain more attention and her/his career will gain from the controversy.
3. There are two general sources of Antifragility: the presence of multiple, low cost elements of the system that can be sacrificed (the cells of the horse, single merchants in a Souk, edits or editors in a Wiki), and/or a high level of optionality (Souk, Bicycle, Freelance…)
4. Fragile entities are either dependent on a single/few variable/s (Head Teacher needs a good reputation) or highly specialised, hence they either have little optionality or they rely on some element of the system that can’t be repaired or replaced.
Slide 6 (Optionality makes science antifragile): [Animated]
In Science, the main source of antifragility is optionality . Because science can proceed by (small scale) experiments, the advantages of what works can be significantly higher than the cost of what doesn’t. Creating an asymmetric, convex payoff curve.
Starting from a symmetric payoff, even when one moves along the X axis by pure chance, if one can discard (part of) the negative payoffs, the result will be a convex curve that can lead to the accumulation of gain.
Slide 7 (Knowledge-based research under high uncertainty):
The original essay includes the “convexity bias” figure described as follows:
The Antifragility Edge (Convexity Bias). A random simulation shows the difference between a) the process with convex trial and error (antifragile) b) a process of pure knowledge devoid of convex tinkering (knowledge based), c) the process of nonconvex trial and error; where errors are equal in harm and gains (pure chance). As we can see there are domains in which rational and convex tinkering dwarfs the effect of pure knowledge.
I was unable to find out exactly how this simulation was produced, so I tried something similar myself (next slide)
Slide 8 (Simulation*):
I used three payoff curves and randomly generated 1000 values within the same range. These numbers (x) were used to calculate the three different f(x) and the results summed to the previous values for each curve.
Convex research is a Parable, Pure chance a symmetric straight line, Pure Knowledge is the same line but moved a little upwards, so that a positive outcome has 55% of probability (slightly better than pure chance).
There are many other ways to simulate each approach, but the one chosen does produce the expected outcome: Convex research improves faster than proceeding by strict knowledge (provided that the domain of inquiry is complex and not well understood).
When dealing with complex (always somewhat unpredictable) systems, a tinkering, not knowledge-driven approach can be highly beneficial if each trial is relatively cheap, and the existing knowledge is not enough to produce reliable predictions. This second condition is always true for Social Science, because the level of complexity of the systems that are studied is always high enough. There is always the possibility that some highly significant unexpected event will happen, even if it was never observed before (see the Black Swan concept: http://www.nytimes.com/2007/04/22/books/chapters/0422-1st-tale.html)
Slide 9 (Seven rules of antifragility in research):
The rules above are directly derived from Taleb’s essay, slightly adapted for the occasion.
Slide 10 (Is this it? You knew it already, right?):
The consequences of the Antifragility concept can be surprising:
Central planning is certainly fragile, because it relies on the negation of optionality. Perhaps surprisingly, the same applies to big corporations, that need to be much more rigid than single individuals or small initiatives.
One could argue that the logical consequence is that optimal policy making should facilitate the formation of antifragile systems, something that is very far from the traditional top-down approach.
On the other hand, understanding the sources of antifragility allows to design both research protocols and interventions in a way that maximises their intrinsic antifragility (or reduces fragility).
The above is extremely important especially when the system studied is not well known (almost always, as Social Systems by definition are extremely complex). It is a trade-off between knowledge-driven “rigidity” and antifragile (cheap) trial and error: the point being that in social science the second should be normally preferable (even if it’s counter-intuitive).
The next slide contains 15 open questions, that the audience was asked to discuss: in small groups of 2-4 people, each group could randomly pick one or two questions to tackle. They then had 1-2 minutes to explain their discussion to the other groups.
Slide 11 (Questions for discussion):
All the questions are supposed to allow multiple possible answers, they should encourage the application of antifragility to the many contexts of our research.
Some of the questions focus on important consequences of antifragility that it was not possible to discuss explicitly. See for example q2, q7 and q9.
Q12 was my little inside joke. I was expecting that most of the audience would doubt that it is frequently possible to inject antifragility in a research programme or policy intervention. To demonstrate that that’s not the case, I’ve used this seminar: by including the groups discussion, with random group members and optional questions, I’ve tried to increase antifragility. If one person was unconvinced, another person in the group may be able to tackle the disagreement; playing with different questions (and being allowed to choose one that suits the single individual) maximises the chances that the main concept will be retained; if someone was unable to apply the concept, after trying my answers would have a better chance to hit the target, and so on.
One would think that a seminar has to be fragile, or robust at best. But in fact, that is not the case, and I hope I’ve demonstrated it. The same applies to research and policy making.
*The Excel Worksheet used to generate the simulation is available here.