From the outset, several mathematicians and computer scientists believed that an AI should possess a logical artificial mind, making it predictable and allowing its capabilities and limitations to be proven.
The Logic Theorist was one of the methods showcased at the Dartmouth Conference, having been developed the year prior by Allen Newell and Herbert Simon. Herbert Simon, a political scientist, later received the Nobel Prize for his theory of bounded rationality, which posits that rational individuals generally make satisfactory decisions, though these may not always be optimal. Cliff Shaw was tasked with programming the Logic Theorist.
The Logic Theorist was based on the concept of using search techniques to discover solutions to problems, specifically to find a valid proof connecting a logical hypothesis to a proposition. For instance, consider the hypothesis that if it rains, the pavement gets wet, and if the pavement is wet, it becomes slippery. If we assert that it is raining today, we can deduce that the pavement will be slippery.
Complex ideas and mathematics can be represented in various logical forms. Similar to dynamic programming, the Logic Theorist operated on the premise that a search could identify the correct proof amidst a vast array of alternatives, visualising this landscape as a tree. The hypothesis served as the root, with numerous branches stemming from it, each representing a potential logical statement derivable from the hypothesis. Somewhere within this expansive tree lies the proposition, requiring the Logic Theorist to navigate the tree to locate it.
The challenge was that as the hypothesis grew more complex, the size of the tree expanded exponentially, rendering exhaustive searches impractical, even for a computer. Rather than breaking the problem into smaller components like dynamic programming would, Newell and Simon introduced heuristics—rules of thumb that would trim away less promising branches of the tree. To facilitate this process, their third innovation was a novel programming language designed to process lists of symbols.
This AI excelled, successfully proving 38 of the first 52 theorems from chapter 2 of the renowned foundational mathematics text “Principia Mathematica,” authored by Alfred North Whitehead and Bertrand Russell. One of its proofs was more concise and elegant than the original presented in the book.
Soon, many people became enthusiastic about this AI, as it operated within the realm of logic and was designed to tackle problems in ways similar to humans. It enabled the representation of concepts through symbols, allowing it to reason about the world using these symbols. Symbolic processing emerged as a fundamental aspect of AI and significantly influenced cognitive scientists for many years.
Today, when many flaws of deep learning networks persist despite considerable research and billions of dollars invested, we wonder whether we might return to some principle of symbolic AI to build alternative and hybrid AI architectures that incorporate causality and reasoning.