What is Artificial Intelligence?
There are many definitions of artificial intelligence or AI (a Google search returns 2.1 billion results). One of the funniest and most vague definition is “AI is whatever hasn’t been done yet.”
Merriam-Webster defines AI as “a branch of computer science dealing with the simulation of intelligent behavior in computers; the capability of a machine to imitate intelligent human behavior.” AI is the ability of a computer program or a machine to think and learn. In general use, AI means a machine which mimics human cognition. Machines being able to think and learn seems to be the crux of AI.
The way humans think is through cognition (from the Latin for know or recognize). It is the scientific word for a thought process, the mental action of acquiring knowledge and understanding through thought and experience. The way humans learn is through either observational or associative means. Observational learning is learning by watching others’ behaviors, such as watching a coworker prepare to enter a jobsite by clocking in and reviewing safety procedures. Associative learning is learning by establishing connections between events, e.g. hearing thunder followed by a lightning strike.
Humans make decisions based on thought and learning, making sound or good decisions based on observational reasoning as well as associative patterns. They also make bad decisions that they learn from, making them smarter, so thought process improve.
If a machine can acquire knowledge and understand or recognize it, then it too can start to make informed (and hopefully good) decisions. It is believed that AI is really about a machine being able to make an informed decision that is a sound one. AI is a decision support system (DSS) that helps make better decisions faster than one could have otherwise made.
The Problem with Project Planning Today
One of the hardest challenges in project management is accurately forecasting future outcomes (project completion date, total cost) of very complicated and highly uncertain endeavors (projects)—which is often called “planning.”
As an industry, there have been developments of some tried and trusted techniques such as critical path method to assist in modeling project outcomes. But these models are only as good as the inputs fed into them. Any worthy planning tool today uses CPM as its underlying forecasting engine. And yet, the planners are still left with the onerous task of knowing not only which activities to include in the plan, but worse, what should their durations, cost and even sequence be? CPM does little more than convert durations and sequences of durations into a series of dates. It doesn’t help answer questions such as:
- what scope should they focus on when building a plan;
- what activities should they include;
- what should their durations be;
- what is the true sequence and logic between their activities; and
- what risks or opportunities can they expect to encounter?
If CPM were a complete solution, then there wouldn’t continue to be project cost and schedule overruns. The problem isn’t CPM though. The problem is the inability to accurately model what might happen during project execution because:
- there are a huge number of variables (tasks and sequence) and
- there are a huge number of uncertainties associated with those variables (duration or scope uncertainty).
Schedule risk analysis tools help to show how bad forecasts may be, but they do nothing in terms of determining what the inputs to the schedule should have been in the first place.
This is why AI can massively help project planning. If AI can assist the planner by making suggestions that are sound, then the immense challenge described starts to become surmountable. Added to that, if the planning tool can also start to make better suggestions by observational or associative learning, then they are headed down a seriously valuable and exciting path.
Artificial Intelligence Categories
If the number of AI definitions was daunting, know that Google returns 754,000 results when searching for Types of AI. Sadly, very few of those results return a common set of type definitions.
It is most beneficial to categorize AI into the following three types:
- Artificial Narrow Intelligence. ANI, also referred to as Weak or Applied AI, is a type of AI that specializes in one area. An example of this is IBM’s Deep Blue computer beating a chess master. The machine was programmed to be very good at one thing – playing chess. Apple’s Siri is another example. While programmed to respond to a limited set of questions, she cannot give an informed answer if the question goes beyond her programming. The majority of today’s AI solutions are ANI-based. Given the current state of AI-related technologies, ANI-based technologies are the most likely to support advancing the science of planning.
- Artificial General Intelligence. AGI is also referred to as Strong AI or Human AI. AGI refers to a computer that is as smart as a human across multiple domains. Computer science is nowhere close to achieving AGI yet.
- Artificial Superintelligence. ASI is “an intellect that is much smarter than the best human brains in practically every field, including scientific creativity, general wisdom and social skills.” This is an even greater leap than AGI.
Current Approaches to Artificial Narrow Intelligence
Today, the implementation of ANI-based AI can be loosely classified into two categories: expert systems and neural networks.
Expert (Knowledge-Based) Systems
Originally developed for use in the 1980s, expert (or knowledge-based) systems (ES) came into their own as computing power got strong enough in the 1990s. An expert system is a program running on a computer that uses a set of rules to answer a question (typically in the form of IF…THEN).
When asked a question, an ES will filter a set of data, based on rules, to establish a subset of what it believes is the answer. In general, the more rules that can be used to answer the question, the stronger the chance that a correct answer will be given. For example, if one wanted to determine a type of two-legged animal, simply querying “IF number of legs = 2” doesn’t narrow down the search enough to give a useful answer since there are a large number of animals with two legs. Combine this with an additional set of questions relating to height, weight, habitat, pouch, etc. and a more reasonable answer can be deduced.
An expert system is comprised of a knowledge base and an inference engine. For a project planning tool, the knowledge base would contain data pertaining to activities and their durations for different types of project. The inference engine is then responsible for trying to return a subset of this knowledge base back to the planner based on the question they may ask, such as “What activities should be included for the engineering scope of a hospital project?”
In this example, what is further useful is to understand how confident is the computer that the returned suggestion is correct. This is where concepts like fuzzy logic come into play. Rather than returning a definitive list of activities, the AI engine should return a subset of activities with associated degrees of confidence about their relevance.
Neural Networks
A neural network tries to simulate the way a brain processes, learns and remembers information. Learning from experience it looks for similarities in information that it is provided, as well as in previous data and then makes a decision based on that process—it looks for patterns. This pattern matching is called machine learning—the NN must be taught what is a match and what is not. Feed in enough examples of characteristics of a living human (breathing, pulse, eye movement) and a neural network will start to establish a pattern as to whether those inputs drive towards a correct diagnosis of “alive or dead?”
There are various forms of machine learning in a neural network including:
- Supervised. For example, feed in an activity that has zero total float and tell the NN that the activity is on the “critical” path. After feeding in enough of these activities, it will establish a pattern that matches zero float activities to critical path activities.
- Unsupervised. Feed in activities but don’t tell the NN which are on the critical path or not and let the NN try and categorize the activities based on its various attributes (e.g., total float). In this instance, the NN will perhaps group into zero and non-zero float without knowing this relates to critical path – it simply groups activities together.
- Reinforcement. This is teaching through reward, like teaching a dog good behavior by offering a treat. With regards to planning software examples, perhaps the license cost of the software should automatically go up or down depending on how good the AI engine suggestions are.
Which AI Approach is Best for Helping with Project Planning?
Unlike neural networks, expert systems do not require up-front learning nor necessarily require large amounts of data to be effective. Expert systems can and do absolutely learn and get smarter over time (by adjusting or adding rules in the inference engine) but they have the benefit of not needing to be “trained up front” to function correctly.
Capturing planning knowledge can be a daunting task and arguably very specific and unique to individual organizations. If all organizations planned to use the same knowledge e.g., standard sub-nets, then they could simply put their heads together as an industry and establish a global “planning bible’ from which everyone could subscribe. This of course isn’t the case and so for a neural network to be effective in helping in project planning, a lot of data would need to mined. Even if it was readily available, it wouldn’t be consistent enough to actually help with pattern recognition.
Neural networks have been described as black boxes—feed in inputs, they establish algorithms based on learned patterns and then spit out an answer. The problem is, they don’t tell why because neural networks don’t understand context. This is problematic, raising the question of whether a diligent planning community should rely on a system that doesn’t have understanding, or even worse, cannot explain why a tool can come to a given answer. A computer providing guidance indicating “You need activities 1, 2 and 3” is not as useful as “You need these activities based on previous projects X, Y, Z and your currently defined scope and the phase of the project you are currently in.” The context is the key to us understanding and ultimately accepting the suggestions made.
Expert systems tend to excel in environments that are more sequential, logical and can be “tamed” by rules—reminiscent of a CPM network. Neural networks pertain more to problems such as recognition through pictures e.g., project drawings and BIM.
Planning can still benefit from a neural network approach as a way to make the tool smarter. As mentioned, expert systems can get smarter, but they need to be trained. If it can track a planner’s reaction to suggestions made by an expert system, then those reactions can be used to potentially adjust the weights given to the various attributes in the expert system.
So then, relating back to the original definition of AI, for a project planning tool, use an expert system to think and use a neural network to learn. Combine these two and the result is an incredibly powerful planning aid.
