On "Why greatness cannot be planned" and how data science happens in the industry

The core concepts from the book and an analysis on how they could be applied in a company using Data Science

Index

  1. Introduction
  2. Core ideas
  3. How the objective problem looks in a company
  4. How much of innovation is expected from a company?
  5. Creating an environment for innovation in a company
  6. Conclusion
  7. References

Introduction

After facing the “Why greatness cannot be planned: the myth of the objective” 1 link for the third time in a Twitter thread about AI research, I finally decided to read it. Though the book brings a general message about innovation, since the authors are AI researches themselves, many examples come from the AI field and there’s a case study about it.

But my main interest was about AI in a company, and that’s the thing I’ve reflected during the reading. Since I’ve felt provoked, uncomfortable and challenged during the read, I feel I had something to learn from it. Here I share the book concepts and my thoughts.

Core ideas

The myth of the objective

Objectives are harmful, especially when we talk about ambitious goals. The authors do accepted the usefulness for simple or near to be achieved goals - things that are one step away.

The book brings a recurrent example about PicBreeder, a site where users can mix images and the most interesting results come from unplanned paths after different users start from other users’ results.

The main allegory used is about stepping stones in a river. The path to ambitious goals is not clear and visible, therefore we can only go from one stepping stone to the other, and that’s when we start using objectives.

To find Homer, one shouldn't look for Homer. Source: PicBreeder

Objectives supposedly help to give people directions. It’s the success compass. You define a couple of objective metrics and then you can monitor them to know if you’re progressing.

The argument is that it’s a false compass. A breakthrough is not preceded by something as important or similar. It includes a disruption. Though cars are faster than horses, cars are not faster horses, which is a classical example.

The authors, Lehman (2011) 2, did AI research on this by changing the objective function of an agent to look for novelty instead of a metric to how close it is from the objective. In the case of an agent in a small maze trying to find its way out, the results show how the novelty search could improve the performance of achieving a goal by not directly optimizing to achieving it. In more complex cases the novelty search struggle, but the argument is that if objective based agents can’t win in a simple problem, what is its chance in a complex one?

Because of these pieces of evidence, objective is framed as a discovery killer, a false guide to great ambitions, something to avoid at all cost, in which consists the book’s paradox: great achievements become less likely when we define objective, to make it more likely to do it, it’s needed to not define it.

Be a treasury hunter

The escape from the previous scenario is to be a treasury hunter. First, we accept that there’s no single search algorithm that can find what you want everytime, following Wolpert’s (1995) 3 no free lunch theorem.

And the algorithm for treasury hunt is to follow novelty and interestingness. The past is a good guide for novelty because we know how things were before. Novelty then can be a guide for interestingness, since we are expanding regions people have never been and it can lead to unknown things. So it means potential. Deviate from the past.

For the treasure hunt consensus can be harmful. The way to go is by separating people from each other, in a way people only interact by taking off from where someone else left. While many might have their own objective, the system as a whole end up lacking a unified objective because people’s objectives differ. One should provide stepping stones to the others.

We should search for everything and anything that’s valuable without looking for something in particular.

Not knowing where we’re going is the way of the information accumulation, it’s the treasury hunter, it’s the stepping stones collector, it’s the path to everywhere and nowhere, it’s the tunnel to the future.

Stanley, Kenneth O., Lehman, Joel; Why greatness cannot be planned

And this is hard!

The path to innovation is not a naive from worse to better. You need to be prepared to experience other ordering. And novelty depends on what we have seen before, it doesn’t progress from bad to good, but from simple to Complex.

How the objective problem looks in a company

From now on, it’s about my thoughts on how Data Science in a company relates to the book conceps.

Objectives would work for extremely short-term and visible goals, but this can be horrible! Companies are accustomed to defining Objective Key Results (OKRs) to guide its team to the right direction, and I believe this is one of the most harmful things you can do with a team if not done properly.

Here’s an example: let’s say our OKR is to grow 50% this year.

If the company is already kind of growing, but when the team assume the current growth rate it’d only get them reasonably below the objective, the natural thing to do is to improve the current process, make incremental changes.

Do you have an ML model in this process? Make a better next version. Do you do marketing campaigns? Make them more effective, improve conversion. Can we tweak a policy?

Back to the cars and horses example, an OKR of a 10% faster transport would most likely make people focus on how to make horses 10% better than trying to find a completely different way of doing it.

And this is the trap that make Data Science and Machine Learning people always prioritize the next version of a model involved in an important OKR.

Notice that having an impossible OKR is not a solution! Because it’s exactly the worse case: an ambitious goal. It’s still misleading, the team will go for incremental changes expecting to move the metric a little, imagining it’s just a matter of accumulating similar improvements to achieve it.

As highlighted by the authors, one of the main reasons we end up sticking to objectives in general is a fear of risk. Those allocating resources don’t want so much risk so resources get wasted in incremental projects.

How much of innovation is expected from a company?

For Stanley and Lehman, in a company, innovation might mean revealing a nearly stepping stone no one realized it was there. But if it’s a nearby thing, it’s more likely other players will follow the same way at pretty much the same time.

It’s easy to see that companies can’t keep collecting stepping stones indefinitely. In fact, it sounds like a cool definition for Startups. But by not collecting any at all, or not being able to communicating them from small groups to other groups inside the company, it’s just waiting to a new player disrupt it and get pushed out of the market.

But instead of proposing a balance between objectives and novelty, which is something the authors argue it’s very natural to try when resisting dropping the objectives, we can think about how to do the minimum necessary on short-term objectives to enable ambitious achievements.

Creating an environment for innovation in a company

Considering the book concepts and a company context, here’s the list of things we’d need:

The reasoning is that to make a Data Science team to become a treasury hunter, you need to free them from short-term optimization and reward risk taking by enforcing that collecting interesting results about data science in the company’s context is part of their job, just as offering it to the whole team as a new path to take.

They key part of making a case for it with leadership is the company’s context and what it has done and how it does. Interestingness is not defined solely by a Data Scientist’s desire to explore a subject they love. If a company has never used computer vision for a particular problem and there’s image data, it means novelty in the company context. If a company is obsessed with customer satisfaction and there’s data about it, so it’s a clue interestingness is lying there and waiting to be explored. It’s not a random search.

I remember someone in my team taking ideas from a model built in a different area. It’s great, but it’s also the minimum expected: someone presents something that works, you take that and apply to your context. At this level, everything is very predictable. What is way more interesting is building a path from multiple teams’ interesting results and being able to get to a point completely unpredictable without their contributions. Communication becomes vital.

My suggestions are purely investigative, personal and things I’d love to try, and I hope I can experience some of them and write about this subject again. In the following sections I detail my vision about the requirements to become a treasury hunter.

Clear business constraints

People will naturally have the constraints in their minds when looking to the most interesting path. If you’re working with fraud, you know you can’t let too many cases to happen, otherwise you’re out of business.

So instead of having metrics as objectives via OKRs, a team must have a few constraints to keep the business profitable. Which means stop pushing to improve the metric, like: “Prevent 90% of the fraudsters. Now 92%, then 94%…”, and just define the minimum acceptable value: “we need to operate blocking at least 90% of the fraud cases”.

It doesn’t mean you can’t have a list of important metrics you monitor daily or weekly. You will keep it. It’s just not your goal to move it, though you’ll if you succeed on the treasury hunt.

Though the authors tell us we don’t need to constraint to avoid meaningless options, that the impossibility of taking most of the options is enough to make the stepping stones behaviors that actually work, here I’m considering constraints the “business impossibilities” that might not be clear to everybody, like “we can’t survive with this cash flow”.

A good understand of the past and the present

To be able to compare past and present to a next step to evaluate how novel and interesting it is, people need to understand past and present. And this is sometimes hard in a company.

In academia, it’s part of the PhD training to review the bibliography, but in a company people will have at most the onboard in a brief period with high level information.

So Documentation, accessibility and an effort to make the company’s journey clear and show how was the process of collecting the stepping stones to get to where you are is essential.

As a data scientist, I enjoy and enforce the creation of different artifacts during model creation. Research phase is also documented with short takeaways. So there’s something to point to when you onboard people and it avoids the “oh, we have tried it in the past!” when a new person suggest an idea. This situation happens because they don’t have yet a sense of what is novel and interesting in your context.

Educate people to know what interestingness means

It’s expected you’re hiring Data Scientists and Machine Learning Engineers with great knowledge to know what is novelty in ML. So hiring well and educating people to a point they can see other teams’ presentations and identify the stepping stones they are offering to them.

But then there’s the business side of it. How could someone knows that doing a certain process with a certain model had and effect X and X is interesting? I think here we get really close to the trap of defining an objective. I’d set this as a general vision, principles, or values.

So if customer satisfaction is a company core value, things relate to it get more interesting. Remember what was considered an interesting outcome for PicBreeder: images about things humans consider interesting. So in a company it’s the same thing, the sense about interesting should come from : things this company considers interesting.

At the same time, there was no particular objective on PicBreeder, and the same thing goes for company: it shouldn’t define it.

Make everyone’s stepping stones visible to everybody

If there’s something valuable in having different areas, groups and people looking to different aspects of a business is the divergence search it provides to the whole system. But then it needs to be communicated.

It’s needed to have an efficient system to share interesting results. And this is completely different from the company wide communications talking about the current stage of the OKRs, on going projects stages, and deployed solutions, because any of those things is offering stepping stones, they are executive summaries.

It’s for sure a challenge to communicate in the right level and don’t overwhelm. But if a group get attained to novelty and interestingness, they can filter their communication to focus on this. Think about your last project. Now try to list the surprising results and novel stuff from it. There’ll be only a few, it’s a way shorter thing to spread than communicating the whole project.

Considering the teams I’ve been working, we both influenced others and took certain paths influenced by different teams. It’s very common DS people can identify it during an 1-2h presentation, but it won’t scale. You can’t attend many 1h project presentations hopping to get something new.

It needs to be something like a novelty-tweet thing.

To make this system work, you need to equip the team with the previous step of giving them a sense of what is interesting, otherwise they will only share results able to move their particular metric, and many stepping stones will never be part of the whole team possible paths.

It takes a lot of culture to make a team confident enough to share something they have worked for a while and the main result is interestingness considering the narrow view would consider it useless because it’s not optimizing a metric.

Make room for disagreement

Consensus was stated as harmful, and if there’s something management looks in a company is consensus from stakeholders by informing and giving visibility on their teams objectives. So we don’t just define them, but we make them converge to upper management view of importance.

So a company should be made of controversy. You shouldn’t feel comfortable with everything your direct reports are doing, but at least a few people in the company are! And then you should let it go this way instead of negotiating something in the middle.

The same thing goes about your discovery path and your manager, until you get to the point you shouldn’t have consensus even from the company board.

No uniform performance assessment

Performance management should balance stepping stones collections and impact. If we already have the common behavior of the resources’ owner not wanting to take risks, and we combine to motivating everyone else in the company to not risk also, then the whole thing is doomed.

Drive impact is easier following objectives, short-term, or even… nearby innovation, which is desired.

But if we don’t incentive stepping stones collections, how could anyone reach nearby innovation by use its own or getting from the others?

If you balance novelty and impact, then we could at least collect a few new interesting facts, while delivering some short-term stuff or executing innovation enabled by previous stepping stones.

Enforcing uniform assessment will make people converge, it undermines future discovery, people can’t explore new ways.

It’s needed to give people back their intuition, autonomy, and creativity.

Performance Management is not the problem, but using a standard metric for it to pursue ambitious objectives. A good way of doing it is to assess where you are in respect to where you were instead of where you want to be.

Conclusion

It was certainly an interesting reading! I couldn’t stop thinking about how I’d change things in my work environment to give it a chance. From the suggested things, some are hard to imagine in place, others are closer to local decisions I’d be very open to try.

References

Preview image by @yer_a_wizard.

  1. Stanley, K. O., & Lehman, J., Why greatness cannot be planned: the myth of the objective (2015), : Springer. 

  2. Lehman, J., & Stanley, K. O., Abandoning objectives: evolution through the search for novelty alone, Evolutionary computation, 19(2), 189–223 (2011). 

  3. Wolpert, D. H., Macready, W. G., & others, , No free lunch theorems for search (1995).