Part 2 of 3, By A M Howcroft, assisted by ChatGPT

In part 1 we looked at how problem-solving is an art but would be far more effective if it gained some scientific rigor. We also looked at the critical nature of problem-solving to show it’s not a nice-to-have, but a survival skill for people and organizations. Let’s now take a deeper dive into discussing tools that have been developed to help people structure and solve problems and see why they are not mainstream. If they were highly effective, we would all be using them, right?

The problem with problems.

We should start by tackling the elephant in the room; let me just say it, not all problems are the same. It may therefore help if we start by describing problems and their key attributes, so we can correctly categorize them. Consider the few problems below:

My child has cut her finger, where are the bandages?
How do we solve world hunger?
Which facilities should we open and/or consolidate to maximize our profit?
What is the best car for me to buy next?
What is the general proof for the Goldbach Conjecture?
My parachute has not opened…what do I do now?
What is the most profitable way to allocate inventory to customer demand?

There are some clear differences between these examples, as they vary from personal to global in scale, some require deep expertise, and there are obvious differences in the timescales in which a solution is required and in the necessary time it may take to achieve a solution. There are also issues of quantifiable vs quantitative results; a mathematical proof for Goldbach is something we can test, but how do we measure the ‘best’ car, which clearly depends on subjective requirements and constraints. We can also break the questions into those that have a known solution (there is a box of bandages in your house), to those which have no commonly agreed resolution.

We need a Swiss army knife.

One of the big lessons to come from AI research on Machine Learning (ML) is that there are many different approaches to learning, whether you’re a human or a machine. There is a veritable Swiss army knife of learning options, with each tool being good for certain tasks. In ML, this includes techniques like decision-trees, Vector Machines, and Bayesian models. As human school children many of us learnt the times tables by rote: two twos are four, three twos are six…, and we discovered many other life lessons via heuristics (aka trial and error). Each approach has value, and modern ML looks at how to combine various approaches with the critical trick being how to decide which one best suits any given situation.

This Swiss army knife or multi-tool approach is equally applicable to problem-solving. Just as we use different learning techniques, we can and should use different tools for problem-solving. For that to be effective though, we need to find better ways to categorize our problems, to help us select which tools to use.

A rough framework.

Here’s a framework to consider for problem categorization. This is a simplified version of a more advanced model we developed for our digital agent AVA at SWARM, where we have broken down the primary considerations into three dimensions: the type of problem, its complexity, and the timeframe. The bottom section of the diagram illustrates how problems of each combination are solved by most people and organizations today:

We can easily add several other dimensions to this chart (although it becomes much harder to read) like industry, known/unknown, quantifiable vs quantitative, and so on. We could also expand the existing Type layer to add Global, rather than just Personal or Organization, and we could spread the Difficulty category into Simple/Medium/Complex, and add a long-term Timeframe, too.

The white solution boxes that appear in the bottom row, can also be expanded to describe the best techniques to be applied, and these can be further broken down by other factors, e.g. industry. Ultimately, we can construct a chart (or algorithm) to help us match problems to the ideal problem-solving approach.

Many good tools.

Inside those white boxes, we can find many useful tools for problem-solving. If we take an example from the organization/complex/mid-term silo, we might wish to describe a problem about recurring manufacturing defects. This may be analyzed by a team using causal analysis software tools, or with the paper-based five whys approach which aims to understand the root cause of a problem. Here’s a brief example of five whys in action:

Problem: Why can’t our aircraft take off from the airport?

1. The starboard engine is not delivering enough power.

2. [Why?] the fuel line is not delivering the correct quantity of fuel.

3. [Why?] there’s a blockage in the fuel pipe.

4. [Why?] the fuel is not of the right quality and contains sediment.

5. [Why?] the aircraft was refueled from the wrong tanker.

We can keep asking further why questions to find out if the pilot was new to the airfield, or if a junior operator made a mistake and sent the wrong tanker to the plane, but the intention is to get to the root cause of the problem. The technique is sometimes combined with Ishikawa diagrams (aka fishbone charts), and there are even specialized versions for Manufacturing, Service industries, and so on.

With five whys it’s critical to take the questioning far enough to reach a root cause and not the symptom (imagine if we had stopped at three whys, above). There are also many scenarios where the root cause is a combination of factors, e.g. Why did company X fail? Tools like five-whys are great at what they do, but if you consider our categorization chart above, it’s clear they are not suitable for all scenarios – especially for short-term problems.

Sledgehammer to crack a nut.

Being such inventive apes, we have devised many different problem-solving tools, which can be very powerful. One of the most common reasons a problem-solving attempt fails is because the wrong tool was applied to the problem. As the adage goes, you shouldn’t use a sledgehammer to crack a nut. There’s another issue at play here in the broader market, which also has a well-known saying:

To a lumberjack, everything looks like a tree.

Working as a salesperson at a software vendor, selling a small number of applications, it is far too easy to see any problem you encounter as suitable for your tool. There are clear financial incentives to view the world this way, too. While the best salespeople aim to be trusted advisors, recommending the right tool, there are plenty of pressures being applied to them which can distort that vision.

AI (and especially LLM) is a very large sledgehammer.

Although it’s very powerful, AI is simply another tool. If we use it incorrectly, we won’t get the full benefit, and worse, we risk putting our lives or business in danger. Applied correctly, the benefits are gigantic. Let’s take a look at how AI impacts our decision making for different problem categories, by adding an additional row to our earlier diagram:

One of the key takeaways is that AI is going to have a major impact on how we solve many types of challenges, in both our personal and professional lives. Another point is that the natural starting point for AI is in ‘the middle’, because it’s not so useful for simple/short term challenges like ‘Should I run away from that lion?’, and also cannot be left to independently solve complex long-term problems such as ‘How do we prevent global warming?’.

In between, there are many areas where AI can be deployed for significant gain.

Given that AI is really a catch-all term that covers many different algorithms and approaches (ML, image-processing, perceptrons, LLM, and even fine-tuning vs RAG within LLM) then we still need to decide which AI tool to use for which job.

In the final post of this series, we'll explore how to map AI tools to categories of problems, and look at the potential impact this may have as we transition, perhaps very rapidly, into making problem-solving a science instead of an art.

Do you have the right tool to solve your problem?