How Machine Learning made AI forget about Knowledge Representation and Reasoning
A brief history of how I learned about the forgotten core of artificial intelligence.
In my early days working as a data scientist in AI I was taught one thing above all: you need more and better data to feed your learning systems. And I have dedicated myself to finding that data.
I was lured into the world of machine learning while trying to discover the world of artificial intelligence. I admit, it is exhilarating to make a computer do complex things to my liking without saying how it should do this. I ran riot in this world and have long forgotten my original goal whilst studying how to prepare data, engineer features and build deep networks.
However, in some cases where I was asked to automate and optimize a business problem, the issue of the availability, quantity and quality of associated data often remained wanting. In addition, the increased acceptance of artificial intelligence in the industry gave rise to fear, skepticism and resentment. People think they are being replaced by a black box. That is not the truth. But I do understand the resistance when you try to replace years of expert knowledge with an opaque box.
These observations made me think again about what AI really is and what methods we can use without hesitation. And I do confess, I only knew about machine learning.
Artificial intelligence is commonly described as the discipline of computer perception, reasoning, and action.
John McCarthy coined the term ‘artificial intelligence’ in 1955 and organized the Dartmouth Research Project in Summer 1956 — the start of AI as a field.
He stated: “The study is to proceed on the basis of the conjecture that every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it. An attempt will be made to find how to make machines use language, form abstractions and concepts, solve kinds of problems now reserved for humans, and improve themselves.”
He developed Lisp in 1958 and reinvented how we program to create thinking machines. And others followed him, e.g. Alain Colmerauer and Philippe Roussel, who came up with Prolog in 1972. Prolog was partly motivated by the desire to reconcile the use of logic as a declarative knowledge representation language with the procedural representation of knowledge.
Thus, Knowledge representation and reasoning (KRR) became a key area of AI concerned with how knowledge can be represented symbolically and manipulated in an automated way by reasoning programs. In other words, how an agent uses what it knows in deciding what to do.
Consider the task of baking a cake. Of course, I could collect data on the amount of ingredients and the order in which they are usually mixed to achieve the best results. But I like Grandmother’s cake best anyway and I know the recipe very well. So why shouldn’t I explain to the system right away which rules have to be observed when mixing the ingredients and which basic conditions have to be created? The same happens when we study complex processes in industry. Too often, people first reflexively ask for existing data and weigh up what additional data should be collected for a successful project. Only a few shrewd people are interested in the details of the processes underlying the problem. It often turns out that the automatization or optimization is just a combinatorial problem, albeit a very complex one.
From now on, when I think of a new problem, my first question no longer refers to the data, but to the problem itself.
The development of a process, often over many years and enriched with a lot of experience and expert knowledge, makes it increasingly complex. New dependencies are being added and the constantly growing requirements demand more precise and above all faster results. A task that can simply no longer be accomplished by the human brain. With the digital and systematic recording of the existing knowledge as well as the detailed process steps and constraints, KRR can not only automate the process but also optimize it. Examples are work shift planning, radio frequency auction, time tabling and many more.
Thus, learning from data is extremely popular — but not your only option.
With all the knowledge of machine learning and looking back at what has already been achieved in the field of AI, we should always use one of our strongest and most human abilities: creativity. Think outside the box and do not choose a method because it sounds best or attracts attention. Choose the method that best fits your problem.
Bottom line: Do not let a black box naturally decide how the processes are designed. Talk to the experts and map their process. With the necessary transparency we create trust and pave the way for a higher acceptance of artificial intelligence within society.