Traditionally, technology in construction has always been applied to increase efficiency and productivity. These days, machines and analytical systems can go further, providing useful advice on how to plan, schedule and execute projects through the “application of Artificial Intelligence (AI),” which can be deemed as quite a broad statement, when someone starts considering the vast number of applications and technologies under that category.
Artificial intelligence is a broad category that includes a variety of sub-disciplines:
- Recognition - Speech, Visual & Text
- Machine Vision
- Natural language processing (NLP)
- Machine Learning
- Comprehension, e.g. Expert Systems
The above sub-fields of AI use various tools and models, including different versions of search and mathematical optimization algorithms, artificial neural networks, and others e.g. statistics.
So, with little distinction around the goal(s) the AI is meant to achieve or what capabilities it aims to enable, the most popular response (or predominant notion) is probably that AI = Machine Learning (ML). This is somewhat true, but given the increasingly optimistic future of AI applications in the field of construction, it’s important to try and understand it more clearly in order to maximize the gains of the technology.
Machine Learning (ML)
Machine learning is based on the core idea that algorithms can improve automatically with the use of data -- in other words ‘learn through experience.” As the analytical model gets better, it can start making predictions or decisions without being explicitly programmed to do so, by detecting patterns.
Simply put, the “experience” equates to the “training data” that is fed to the algorithm. During “training,” there might be cases where models yield numerous answers and one might need to label certain correct answers as valid. This labelling also serves as “training,” and as the algorithm becomes smarter in detecting the correct answers, it can make accurate predictions and data-driven decisions. On that basis, data collection and training not only are essential but can be long and tough processes as well.
Machine learning includes the following main approaches:
- Supervised learning: machine learns from example inputs and their desired outputs, as defined by a "teaching agent" (see labeling example above). The goal is to learn a general rule that maps inputs to outputs.
- Unsupervised learning: machine learns only from example inputs, i.e. without labels, and discovers input data patterns.
- Deep learning: machine mimics the human brain functions by performing human-like tasks without human interjection e.g. self driving cars. It’s main implementation is done by neural networks and needs tons of data and computational power so that the system can draw conclusions on its own.
- Reinforcement learning: computer program aims to maximize the rewards, provided via feedback, as it performs a certain goal, e.g. playing a game against a virtual opponent.
Machine learning employs many different methods and techniques to teach computers how to accomplish tasks and is closely related to the field of optimization, as many learning problems are formulated as a loss minimization function.
One of the most important - and probably less known - areas of Artificial Intelligence is the search algorithms. Most often, AI software is using some search algorithms in the background in order to achieve the prescribed tasks or solve complex problems and provide the optimal solution. In AI, almost any problem can be outlined using the concepts below:
- State - A potential solution of a problem.
- Transition - The action of moving between states.
- Start State - Where the search starts from.
- Intermediate State - All states between the starting and the goal state which the algorithm needs to transition to.
- Goal State - The state to stop searching.
- Search Space - A collection of possible solutions to a problem.
Searching in AI is a step-by-step process of navigating from a starting state to a goal state by transitioning through intermediate states. Problems are represented by a “Search tree,” the nodes of which correspond to the various states (solutions). Depending on the search objective, a condition is defined to test the satisfaction of the search objective in order to determine the optimal solution has been found. Broadly speaking, the algorithm aims to reach the goal state by investigating the various alternatives in different ways, depending on the type of search strategy employed.
Example of a Search Tree - node 2 is the start state where the search begins.
The main searching strategies are:
- Informed search - there is information or accurate estimates on how far the goal state is, i.e. there is domain knowledge.
- Uninformed search - it’s unknown how close the goal state is, i.e. no domain knowledge.
- Local search - used when there are multiple goal states and the best one needs to be located.
Each of these has its own benefits and drawbacks and solves different kinds of problems.
While technological advancements open possibilities and are (for the most part) super cool, let’s not forget their dependence on high quality digital information. Sometimes this is referred to as the GIGO principle - Garbage In, Garbage Out. Despite AI having transformed many fields, it has often failed to deliver the expected results due to a variety of reasons, such as absence of suitable data, data bias, wrong people and many others.
Lastly, the underlying complexity of these systems can sometimes be quite challenging to understand, as the algorithms and models can be intricate, to say the least. This limited transparency can potentially make the appropriate adoption and use of AI tools a tough nut to crack. As such, continuous work on enhancing the explainability of the used models will be proven gradually more and more important.
Mitchell, Tom: Machine Learning.
Bishop, C. M: Pattern Recognition and Machine Learning, Springer,
Biere; M. Heule; H. van Maaren: Handbook of Satisfiability
Russell, Stuart J.; Norvig, Peter: Artificial Intelligence: A Modern Approach