1

Introduction

This chapter will focus on the initial conception of artificial intelligence, the beginning stages of its creation, and its development into what it has become today. Even before computers were invented, people have been obsessed with the idea that a machine could think for itself. The Turk machine in the late 18th century was a famous attempt at fooling people into believing the first intelligent machine had been created. Once computer technology advanced, people such as Alan Turing saw a future where creating a truly intelligent machine might be possible. Further development of computer hardware and software then led to heuristic problem solving, a crucial step in creating artificial intelligence. Finally, developers were able to implement something that Turing had theorized years before, the neural network. These days, we are able to create AIs in multiple ways with a variety of techniques for training and implementing them. They can be created on almost all platforms of modern computers and are used throughout scientific fields. With artificial intelligence being so widespread and constantly evolving, it’s important that we take a look at how it all began. By understanding the conception of an AI by Alan Turing, the development of heuristic problem solving, and the theory behind neural networks, we can ensure that citizens are capable of staying informed about current and future policy changes.

“Laptop Brain” by Mohamed Hassan is in the Public Domain, CC0

Alan Turing

Alan Turing typically receives the credit for the initial theory and conceptualization of artificial intelligence, and rightfully so. Turing was an English computer developer in the 1900s who created many logical and theoretical models that were crucial in the advancement of computing technology. He was the first one to describe a Turing machine, which is a system capable of executing and modifying instructions stored in programmable memory. He gained worldwide fame for his creation of the Enigma machine that was used to break the German encryption codes used in World War II. In the late 1940s, he began to explore the ideas of artificial intelligence (Copeland, 2020).

In a report by Alan Turing called Intelligent Machinery, he discussed computers that would be able to learn and think. He often applied these concepts to the game of chess. Chess lends itself well to artificial intelligence because its basic principle and rules are relatively simple. But creating a chess AI is not as simple as calculating all of the possible moves that can be made in a given position. This is because the number of moves that can be made grows exponentially as the game progresses. Calculating every combination would take an unbelievable amount of time for each move that is made. Even today, chess engines have maximum depths that they can calculate to, typically 10-25 moves ahead, depending on the position. And even then, they don’t have an immediate answer to every possible move the other player makes. Sometimes they require time to think in order to respond to a move they didn’t calculate. This will make more sense once we talk about heuristic problem solving.

The Turing Test

“Turing Test Version 1” by Hugo Férée is licensed under CC BY-SA 3.0

The Turing Test is an exercise created by Alan Turing based on the Victorian Imitation Game. In this game, there is a man, a woman, and an interrogator. The interrogator asks questions in order to determine which person is which gender. The man tries to deceive the interrogator in determining the identities, and the woman tries to help the interrogator. In the Turing Test, the man is replaced with a computer. In this scenario, the interrogator must determine who is the computer and who is the human (Saygin, 2000). Hugh Loebner, a businessman, even set up a competition for $100,000 that would be given to the first AI capable of consistently fooling a majority of interrogators. Unfortunately, no AI has even come close to winning yet; however, this doesn’t mean artificial intelligence hasn’t made incredible strides over the years.

Heuristic Problem Solving

If you were to sit down and try to create an AI that could beat anyone in chess, how would you do it? You might think that it’s easy, all you have to do is look at every possible move then choose the best one. Sure it might work for one move, but what if you’re trying to look multiple moves into the future? That’s when you start to realize that the more moves into the future you look at, the amount of moves you have to consider grows exponentially. If you told the computer to calculate every move until the final one in the game, you might be surprised to learn that it would take longer to compute than the entire length of time that the universe has existed. This is the problem with traditional computing. As humans we tend to have computers take the brute force approach. But this isn’t how we would go about playing a game of chess, so why would we make the computer work that way?

Heuristic problem solving is the process in which decisions are guided through information given by memory or learning in heuristics. A heuristic can be thought of as a shortcut that cuts down on the number of calculations that need to happen in an algorithm (Fox, 2020). Humans do this all the time when solving problems, so it would make sense for computers to think the same way. In the chess example, rather than looking at every single move, what if the AI disregards moves which are obviously bad, such as giving away a piece for free? There’s no need to calculate every position after that move because it will objectively be a worse position than the other options. Although the computer may not achieve the perfect answer, it will reach an acceptable one incredibly faster. Once computers had advanced well beyond what they could do in Alan Turing’s time, developers interested in artificial intelligence began implementing machines capable of making these guided decisions. In order for it to be possible, they had to have enough memory to store previous knowledge and mechanisms to edit that memory when new information was learned. This was first implemented in programs designed specifically for individual computers, such as the EDSAC and IBM 701 (Copeland, 2020). Eventually, languages were developed which were able to create artificial intelligence programs on multiple computers.

Neural Networks

“MultiLayerNeuralNetworkBigger” by Chrislb is licensed under CC BY-SA 3.0

Once heuristic problem solving was possible on computers, people began to experiment with another concept outlined by Alan Turing: neural networks. These consist of individual neurons, similar to a brain, that are connected using weights (Rafiq, 2001). Depending on the weights of these connections, a network would be able to produce a set of outputs based on a set of given inputs. A simple example of a neural network could be given a single input, which is a grayscale picture of a handwritten number. The outputs for the network would be the numbers 0 through 9. Based on the weights of the connections, the output that has the highest calculated value would be determined to be the correct number. This concept is very hard to understand at first, but the inner workings of neural networks are actually inherently mysterious.

Neural networks contain one or more “hidden” layers of neurons that are connected to every neuron in the layers on either side of them. When the network is initialized, the weights of the connections are randomized from -1.0 to 1.0. A positive weight indicates the connection as excitatory, and negative as inhibitory (Song, 2016). At each layer, all inputs to a neuron are multiplied by the weight of their connection and summed to produce the value of the neuron. This then continues through the network until the output layer is reached.

AIs are typically created beginning with random weights, which predictably produce random outputs. This is why the neural network must be “trained” through a supervisor tuning the weights using high-level calculus in order to achieve the desired result. This training requires large amounts of data for the neural network to learn from, which is one of the major challenges that artificial intelligence developers face. One common misconception about AIs is that they always make the correct decision. This is actually incredibly wrong once you understand the methodology behind how neural networks operate. Instead of providing answers, the network produces its best guess based on the weights that it has been tuned to respond to. AIs are typically only implemented after extensive training and tuning, usually until they can provide above a 99% success rate. This means that they are not as “all-knowing” as people might think. They are almost always designed for only one task and aren’t even perfect at doing it.

Recent Developments

Now that artificial intelligence has received more focus and implementation, people have started developing new methods and practices for creating and training them. One of these developments is the use of an AI to train a new neural network without the need for a supervisor. This means that AIs are already learning from each other in order to create larger, more complex networks. Another recent development is evolutionary development through genetic algorithms. In this kind of training, multiple neural networks are created with the same structure but different randomized weights and grouped into a generation. They are then run through a trial and the one which performed the best is selected as the survivor of that generation, and modified copies are created until a new survivor beats the previous one. Similar to biology, genetic algorithms use survival of the fittest and evolution to find unique solutions to problems and create the best possible AI.

References

Copeland, B. J. (2020). Artificial intelligence. Britannica. Retrieved November 23, 2020, from https://www.britannica.com/technology/artificial-intelligence

Fox, P. (2020). Heuristics & approximate solutions. Khan Academy. Retrieved November 23, 2020, from https://www.khanacademy.org/computing/ap-computer-science-principles/algorithms-101/solving-hard-problems/a/using-heuristics

Rafiq, M. Y. (2001). Neural network design for engineering applications. Computers & Structures. Retrieved November 23, 2020, from https://doi.org/10.1016/S0045-7949(01)00039-6

Saygin, A. (2000). Turing Test: 50 Years Later. Minds and Machines. Retrieved November 23, 2020, from https://link.springer.com/article/10.1023/A:1011288000451

Song, H. (2016). Training Excitatory-Inhibitory Recurrent Neural Networks for Cognitive Tasks: A Simple and Flexible Framework. PLos Comput Biol. Retrieved November 23, 2020, from https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4771709/