It is Artificial, but is it Intelligent?
Ty Newell, PE, PhD
Vice President Build Equinox
Emeritus Professor of Mechanical Engineering, University of Illinois
Artificial Intelligence, or “AI”, is promoted in every direction one turns. The New York Times (AI Start-Ups Face a Rough Financial Reality Check, April 29, 2024) reports that over $300 billion has been invested in 26,000 AI firms in the past 3 years, or nearly $1000 for each person in the US! CEOs scramble to incorporate AI terminology into anything and everything, regardless of need and effectiveness. Sales personnel are similarly touting their company’s use of AI, but when asked why it’s important, one often hears crickets.
Continuing advances in computing power and data storage allow more and more powerful AI algorithms to be developed with applications ranging from autonomous driving to medical diagnoses to generation of paintings, music and stories. As with many things, AI can be used for good and bad.
Kurt Vonnegut’s “Player Piano” novel from the 1950s explored a world in which automation eliminates the need for human labor. Vonnegut’s world, the technical equivalent of a tropical paradise, devolves into revolution by dissatisfied, purposeless humans.
HAL, the wayward computer in 2001: A Space Odessey, was born in Urbana Illinois, as was Build Equinox. Arthur C Clarke and Stanley Kubricks’ novel and film envisioned a future with error-prone, human conversing computers. HAL’s “intelligence” decides it needs to sacrifice humans to save itself as HAL senses its human companions plotting to disconnect HAL’s control of their spacecraft.
What is AI and what can it do?
Why is AI important, and not so important for maintaining healthy indoor environments?
Build Equinox and AI
AI has a longer history than many realize. Other names, such as “fuzzy logic”, “expert systems”, “knowledge-based systems”, “machine learning”, “image recognition”, “pattern recognition”, and more are categories of AI. In general, an AI system is something that can be trained to learn about something, and then be used to predict, identify, and/or control other things based on what it has learned.
Among the questions AI researchers seek to answer are:
1) How much “training” is required?
2) How much (computer) time is required to make a decision?
3) How accurate is the AI decision making algorithm?
In my research at the University of Illinois, we used AI in the 1980s to classify containers for an automated recycling plant sorting system. Figure 1 is from one of our publications on our AI sorting research. Sorting containers is a classification problem, similar to medical diagnoses viewing MRI scans. Multiple types of containers move through recycling plants, such as clear glass, green glass, brown glass, clear plastics, color plastics, aluminum cans and steel cans.
The simple human act of sorting containers is very complex. When a human sees a baby food jar, even if it is covered by a label and has smashed peas or carrots coating the jar, people can easily and quickly identify it as clear glass. Similarly, the difference between a green wine bottle (glass) and a green soda bottle (PET plastic) is easy for a human to discern without conscious thought.
More than 30 years ago, computer data storage and processing time were limited in comparison to today. For our recycling project, we would collect approximately 200 data points per container consisting of light transmission and reflection data, and acoustic frequency data. We would use groups of 10 to 1000 random containers for training our AI learning algorithms. Once algorithms were trained, other container batches, different from the training containers, were sent through the sensor system for classification. Note that multiple algorithms can be trained simultaneously.
Some AI algorithms, such as “neural networks” compact data characteristics into computationally fast and efficient equation relationships by tuning equation parameters. Just like slope and intercept parameters define a straight line, and choosing a value for “x” determines the value of “y”, multiple parameters are tuned for the neural network equations to develop a model that predicts output values.
“Instance based” AI algorithms, in contrast to neural networks, store all example training data. After training, an analysis routine determines which set of stored example data is mathematically closest to the new input data.
For our container recycling classification project, a neural network AI algorithm required training with 800 containers in order to achieve the same accuracy (30 to 35% error) as the instance-based algorithm trained on 20 containers. Training the two types of algorithms with a set of 10,000 containers reduced the neural network error to 30% while the instance-based algorithm’s error was reduced to 10%.
Although reducing neural network error requires thousands of training examples for acceptable accuracy, its decision-making speed is much faster than the instance-based algorithm. For example, after training both algorithms with 1000 examples, the neural network could classify containers in less than 10 milliseconds while the instance-based classification required 10,000 milliseconds (10 seconds) to perform its computationally intensive comparison analyses to reach a decision. A human requires about 1 second per container for decision making and sorting, with much higher accuracy than either algorithm at that time.
Just as we have seen tremendous advances in renewable energy, flat screen TVs, online shopping, cell phones and almost everything, AI has changed tremendously over the past few decades. Computers have unbelievable operating speeds with nearly unlimited memory. Now, instead of a few hundred information inputs for training, the amount of training data has expanded by multiple orders of magnitude while also reducing decision making speed.
Tesla has employed more than 1000 “data labelers” to identify objects in video streamed from Teslas driving around the world. Human labelers endlessly identify objects in video frames to train Tesla’s AI self-driving software. In addition, auto-labeling algorithms continue to advance, increasing training speeds and reducing the need for human labelers. Every time someone drives a Tesla (other car manufacturers are doing this, too), your driving actions relative to the car’s video and sensors are also used to train algorithms.
Tesla temporarily opened its full self-driving option to all vehicle owners. It was a nice gesture allowing Tesla drivers to consider purchasing the full self-driving mode option. Equally important, it was an opportunity for Tesla to collect a large set of self-driving data for their AI system. Deb and I didn’t purchase full self-driving as we felt it was like riding with a teenager with a learner’s permit. We experienced the thrill and stress of our four kids learning to drive, and have no desire to repeat the experience for Tesla’s benefit.
Finally, while AI researchers understand how their various algorithms “learn”, they do not understand the reasoning behind AI algorithm decisions. We do not know why it has made a good decision or why it has made a bad decision. Some bird crap on a stop sign, similar to researchers confusing a car by placing pieces of tape on a stop sign, may be enough to cause a horrendous accident. As AI is primarily an interpolative decision maker, using AI to extrapolate outside the range of training data should be done with a lot of caution!
AI and Home Air Quality and Comfort Conditioning Control
Build Equinox has been exploring AI research for several years preceding today’s race to implement AI. Readers who have followed our articles and webinars for some time may recall seeing a slide on “Preventilation”, our vision for indoor environments that “sense” what is needed to keep us healthy individually and collectively.
Our Preventilation concept should be viewed two ways:
1) “Preventilation”, management of the indoor environment to prevent conditions harmful to our health and well being.
2) “Preventilation”, a prescient capability to sense and predict the need to prepare an indoor environment for predicted future events.
Figure 2 shows “Big Data – AI” as an aspect of Preventilation (note: this presentation also includes a slide showing how airborne disease transmission is related to ventilation rates…yes, Covid’s destruction could have been significantly reduced if Build Equinox’s air quality recommendations were followed).
Figure 3 shows the heading of a 2016 report written by a University of Illinois engineering student working on an AI research project with Build Equinox. We were examining “AI engines” available at that time to see if they could outperform our CERV2 control algorithms to manage indoor air quality based on an array of input data collected by the CERV2’s sensor array.
We continue to see no advantage for using AI control algorithms over conventional control algorithms managing indoor air quality and comfort (temperature and humidity control). A detriment of using AI, as previously mentioned, is that no one…and I mean no one…will know why an AI algorithm makes a particular decision. With traditional control algorithms, we program in the logic and mathematical relationships for control decisions, and can understand events, and adjust as needed. A wayward AI algorithm will lead to finger pointing and frustration.
For example, a conventional control algorithm compares a desired room temperature with a room’s actual temperature. If the temperature difference between desired and actual temperature is 1 degree, the conditioning system would increase capacity to bring room temperature back to desired temperature. If the temperature difference increases to 2 degrees, the comfort conditioning system’s capacity would be increased to the next level, and so on until conditioning capacity is sufficient to minimize the temperature difference between actual and desired. If something goes wrong, we can follow the logic incorporated into the control system along with sensor readings and conditioning system operation to understand the problem.
An AI system “trained” to manage comfort would monitor a variety of parameters. Not all the parameters would appear to have direct relevance to controlling comfort, but in some manner may be found by the AI algorithm to have correlation significance for controlling comfort. Figure 4 shows a cross-correlation of hourly residential energy, comfort (indoor temperature and humidity) and IAQ (carbon dioxide and VOCs) parameters from 13 CERV-smart ventilated Vermod homes monitored for two years…that is, more than one and a half million data points!
We can visually see some interesting inter-relations among parameters in Figure 4. Some are common sense. The thicker the line, the stronger the correlation among two parameters. Blue lines indicate a positive correlation in which a positive change of one parameter is related to a positive change of the other parameter. Red lines indicate a negative correlation in which a positive change on one parameter is related to a decrease of the other parameter. For example, decreasing outdoor temperatures below room temperature result in higher house energy and HVAC energy usage with a thick red line linking cold outdoor temperature and energy. Clothes dryer energy is strongly, positively related to clothes washing machine usage, as one would expect, resulting in a thick blue line.
Notice that water heating energy and kitchen energy usage are both strongly associated with number of people in a house. An AI algorithm trained on this data might use hot water energy and kitchen energy, along with lesser influence of plug loads, clothes washer energy and clothes dryer energy usage to predict the number of home occupants.
Figure 4 shows very weak cross-correlation between people and IAQ (CO2 and VOCs). How can that be? People and their activities are the main source of indoor pollutants. Weak correlation is a hallmark of CERV2 smart ventilation that automatically adjusts and controls pollutant levels regardless of occupancy and occupant. And the CERV2 does this with known control algorithm control logic because we understand the fundamental processes causing indoor pollution.
The Future
When we understand things, such as controlling air quality and comfort, AI only adds complexity and confusion. When complexity and uncertainty surround a problem, AI is an important tool for finding correlations and characteristic trends that can be used to predict and control. In the meantime, don’t be fooled by marketing and sales with “AI” labeling.
Build Equinox will incorporate AI when we find an advantage over conventional controls for managing health and comfort in our indoor environments.
A few final thoughts about AI:
- Is AI truly intelligent, or is it simply a complex way of determining correlation relationships among large sets of data? If an AI algorithm only had access to the same information as Albert Einstein, would an AI algorithm be able develop a theory of relativity, or would it simply make predictions without understanding?
- Humans have “mindfulness” - a soul. AI has no soul; it can only simulate emotions.
- Environmental sage, Professor Aldo Leopold, in an address to the University of Wisconsin College of Engineering 1938, had this to say about rapid technological developments that continues to apply: “our tools are better than we are, and grow better faster than we do. They suffice to crack the atom, to command the tides. But they do not suffice for the oldest task in human history: to live on a piece of land without spoiling it”
