AI Paradigms and Architectures

The oldest AI paradigm, symbolic AI (sometimes called GOFAI, 'good old-fashioned AI'), builds intelligent behavior out of explicit symbolic representations manipulated by formal logical rules — the founding assumption of the Dartmouth-era field that intelligence could be produced by precisely specifying knowledge and reasoning steps, as Russell and Norvig trace through AI's early decades (Russell & Norvig, AIMA, ch. 1). Its clearest embodiment was the expert system, exemplified by DENDRAL's rule-based inference over chemical structure data, where knowledge is stored as human-readable if-then rules and a general inference engine applies them to new cases (Britannica, 'Expert system'). Symbolic systems have a decisive advantage in interpretability — a human can typically read the rule that fired and understand why — but they struggle badly with tasks like perception or language that resist being reduced to a finite set of hand-written rules, since the real world keeps generating edge cases no rule author anticipated.

Connectionism took the opposite bet: instead of hand-specifying rules, build networks of simple, neuron-like units and let the desired behavior emerge from learned connection weights. Its lineage runs from McCulloch and Pitts's 1943 formal-neuron model, through Rosenblatt's 1958 perceptron — the first trainable network capable of adjusting its own weights from examples — to the 1986 backpropagation algorithm that finally made training deep, multi-layer networks practical at scale (McCulloch & Pitts, 1943; Rosenblatt, 1958; Rumelhart, Hinton & Williams, 1986). Connectionist systems excel precisely where symbolic systems struggle — pattern recognition in messy, high-dimensional data like images and text — but at the cost of interpretability: a trained network's millions of weights rarely offer a human-readable explanation for any individual decision. Where symbolic systems fail loudly and legibly — a missing rule simply doesn't fire — connectionist systems tend to fail quietly, producing a confident but wrong answer with no obvious internal signal that anything went wrong.

A third paradigm, probabilistic AI, treats reasoning under uncertainty as its central problem rather than either fixed logical rules or learned weights alone. Judea Pearl's 1988 book on probabilistic reasoning established Bayesian networks — graphical models that represent probabilistic dependencies among variables — as a rigorous framework for combining prior knowledge with uncertain evidence, giving AI systems a principled way to reason when information is incomplete or noisy rather than assuming either perfect knowledge (as classical logic does) or ignoring uncertainty structure entirely (Pearl, Probabilistic Reasoning in Intelligent Systems, 1988). This probabilistic lens now underlies much of modern statistical machine learning, including the loss functions and probabilistic models taught in foundational courses such as Stanford's CS229, which formalize learning itself as a problem of inference under uncertainty (CS229 lecture notes). Unlike the symbolic paradigm's binary true/false rules, a Bayesian network can represent degrees of belief and update them coherently as new evidence arrives, which is precisely the capability needed whenever an agent must act on incomplete or conflicting information rather than a fully specified rule base.

A fourth paradigm looks to biological evolution rather than logic or learned weights for its inspiration. John Holland's 1975 book Adaptation in Natural and Artificial Systems established the genetic algorithm: a population of candidate solutions is repeatedly evaluated by a fitness function, and the fittest candidates are selected, recombined, and mutated to produce the next generation, iterating until a satisfactory solution emerges — an approach MIT Press's summary of the book credits with launching the entire field of genetic algorithms and inspiring the broader area of artificial life (Holland, 1975). Evolutionary computation is especially useful precisely where neither hand-written rules nor gradient-based learning apply cleanly — for example, optimizing a design or a strategy where the relationship between candidate solutions and their quality is not smoothly differentiable — because it searches a space of solutions without requiring an explicit model of how to improve any single one, only a way to score it. That makes evolutionary methods a natural complement to the symbolic and connectionist paradigms above rather than a competitor to them: they are typically reached for when neither writing explicit rules nor computing a gradient is a realistic option.

A fifth paradigm rejected the assumption, shared by symbolic and much connectionist AI alike, that an agent needs an internal world model at all. In his influential 1991 paper 'Intelligence Without Representation,' Rodney Brooks argued that when intelligence is approached incrementally, with strict reliance on interfacing to the real world through perception and action layer by layer, the need for centralized symbolic representation disappears — his subsumption architecture decomposes a robot's control system into layers of simple behaviors (e.g. avoid obstacles, then wander, then explore) that run in parallel and interact directly with the environment, using the world itself as its own best model rather than building and maintaining an internal one (Brooks, 1991). This behaviour-based approach produced robots that operated robustly in real, unstructured office environments where earlier, plan-then-execute symbolic robots had struggled, and it remains foundational to reactive and embodied robotics. Brooks's claim was deliberately provocative against both symbolic and connectionist orthodoxy of the time, and it is best read not as proof that representation is never useful, but as evidence that certain classes of real-time, embodied tasks are solvable without it. Tasks demanding long-horizon planning, such as multi-step logistics, still generally benefit from some form of explicit internal model, which is one reason the paradigm map in this lesson's practice card treats behavioural and symbolic approaches as complementary rather than as strict rivals.

None of these paradigms has proven sufficient on its own, which has driven growing interest in hybrid neuro-symbolic architectures that deliberately combine connectionist learning with symbolic reasoning. A 2024 systematic review characterizes neuro-symbolic AI as the integration of symbolic and sub-symbolic (connectionist) AI, aiming to combine connectionist systems' strength at learning patterns from raw data with symbolic systems' strength at explicit, verifiable reasoning and interpretability — the review finds research activity concentrated in learning-and-inference and knowledge-representation tasks, with explainability and trustworthiness still comparatively under-explored (Neuro-Symbolic AI in 2024: A Systematic Review). This positions neuro-symbolic AI not as a settled consensus architecture but as an active, still-maturing research direction — a reminder that the field's paradigms are best understood as a toolkit of differing assumptions about how to build intelligence, each suited to different problems, rather than as a single settled orthodoxy that has already won. Which architecture is 'best' therefore depends entirely on the task at hand: a system that must both learn from raw sensor data and produce an auditable, rule-checkable justification for a decision is exactly the kind of problem neuro-symbolic approaches are being built to address, rather than a problem any single pure paradigm from this lesson has fully solved on its own.