How Cognitive Psychology Principles Can be Applied to Knowledge Graphs
There are many cognitive psychology principles involved in explaining human behavior. Today some of these principles can be applied to knowledge graphs to utilize advanced reasoning techniques, improve their accuracy with machine learning outputs, and dramatically increase their ability to fulfill mission critical objectives.
An influential principle for human problem solving originates from Allen Newell’s Unified Theories of Cognition and Human Problem Solving. The principle is that human problem solving can always be described as search in problem spaces. A problem space formally consists of states and operators where a state is any configuration of all the objects and the inter-object relationships that are relevant to a problem. When solving a problem you begin in the ‘begin’ state, and you try to get to the ‘end state’ or solution state. You move forward from the begin state to the end state by repeatedly applying operators to intermediate states.
The game of chess is a good example. The begin state is obviously the board that you just set up, the end state is that you checkmated your opponent. In order to get to the end state you have to apply operators, in this case the operators are obviously the valid moves in the chess game.
Problem solving is about choosing the most appropriate operator and any point in time.
For a beginner just learning ‘what’ chess operators are available is a difficult task in itself. But once you know those rules the more difficult task is to make the ‘right’ move. Over time people that play chess learn an enormous amount of rules or patterns that help with making the right or best move given the current configuration of the board.
Two Ways Humans Learn
There are two major ways of learning which operator to apply. The first one is to try each move in your mind and evaluate the soundness of the new configuration of the problem state. You might do this recursively several levels deep and at some point you make a move. The other way of learning is to just do a move in the real world, and then learn the consequences of this move given this particular situation. Both ways of learning are supported.
The principles of human problem solving and learning described by Newell are all couched in the language of graphs and symbolic rules and patterns, and the more recent versions of Newell’s theories also include machine learning from external operations as a learning and feedback mechanism. Modern intelligent knowledge graphs can borrow many principles from Newell’s work to create learning and self modifying systems.
Applying Cognitive Psychology Principles to a Healthcare Knowledge Graph
Let’s consider a knowledge graph that is a digital twin of a hospital. This knowledge graph represents all relevant entities in a hospital, that is, it knows all patients, nurses, doctors, beds and expensive equipment in a hospital environment. Importantly, it even knows the location of every entity at any point in time. The location is known at a resolution of a few centimeters through RFID and other localization techniques.
Recommended: Google AI Team Explain How ML and Semantic Engines Improve Developer Productivity
In this example, the begin state is the beginning of a shift where we have the current configuration of all the entities. The end state is the end of the shift where every patient is visited a sufficient number of times and there is a minimal amount of suffering for both patients and hospital staff. The operators in this example are every move or visit that a nurse and doctor can make from one entity to another.
There are several factors that make scheduling all these ‘operators’ in a fixed plan not trivial. To begin with: it is sometimes hard to predict how long it will take for a doctor to be with a patient, patients might get unexpected emergencies, and to make matters even more complex, patients can make ‘illegal’ moves such visiting the bathroom when they are not fit to do so. The goal of this knowledge graph is to recommend a schedule at the beginning of a shift, in the full realization that unexpected events can completely change the schedule at any point in time.
When creating the initial plan we look at the state of each patient, but also the availability of nurses and doctors and the goal of minimizing suffering. In the planning process we use symbolic rules and constraints (i.e Patient P cannot get out of bed because of two broken legs; nurse N cannot lift patient P out of bed due to being not strong enough; doctor D is not specialized in the disease of patient P etc.)
But as we mentioned above, during the execution of the plan unexpected things happen: patients get worse, doctors spend more time than expected in surgery, etc. And one can learn from these unexpected events both by learning explicit symbolic rules or by using statistical and machine learning techniques. For example: every time when we get in this particular situation and we respond with this action the outcome is negative. With enough samples we can start learning machine learning rules.
AI ML Insights: Biomed Expert Backs AI’s Role in Improving Human Longevity
The Result: A Self-Modifying System
Using cognitive psychology principles in knowledge graphs can create a virtuous cycle between symbolic reasoning and machine learning – producing a self-modifying system. Self-modify systems can be applied in many domains, but they are especially useful when developing digital systems for power generation plants, manufacturing operations, healthcare services, in the automotive industry, and urban planning.
In all these domains symbolic rules will get you very far but ultimately there are always unexpected events that will force you to fine tune with statistics and machine learning techniques.
Top News: Analysis of Data and AI Penetration Across various Functions
Comments are closed.