Symbolic and search-based AI that predates deep learning and still runs underneath much of it: search and planning, knowledge representation and reasoning, evolutionary computation, and robotics.
Evolutionary computation: population-based, derivative-free optimization. Genetic algorithms, evolution strategies, and swarm intelligence (PSO, ACO).
Knowledge representation and reasoning in AI: expert systems, knowledge graphs, logic, Bayesian networks and fuzzy logic, and how they relate to LLMs.
Robotics fundamentals: the sense-plan-act loop, perception and SLAM, control algorithms, path planning, and where machine learning fits in the autonomy stack.
State-space search and planning: adversarial search (minimax, alpha-beta) and constraint satisfaction problems.
Classical AI is the body of techniques built on explicit representations: states and actions that can be searched, facts and rules that can be reasoned over, and objective functions that can be optimised without gradients. These methods dominated AI research from the 1950s through the 1990s and remain in daily use: route planners, schedulers, compilers, configuration tools, game engines, diagnostic systems and robot navigation stacks all rely on them. Their main strengths are that they are predictable, explainable, data-efficient and can provide guarantees such as optimality or completeness that learned models cannot.
Search and planning treats a problem as a graph of states and finds paths through it, using adversarial search for games and constraint satisfaction for scheduling and configuration. Knowledge and reasoning stores what is known as logic, rules, graphs or probabilities and derives new conclusions; expert systems, Bayesian networks and fuzzy logic live here. Evolutionary computation optimises by maintaining a population of candidate solutions and applying selection, crossover and mutation, which works when the objective is noisy or non-differentiable. Robotics integrates all of them with perception and control in a continuous sense-plan-act loop.
| Problem shape | Technique | Example |
|---|---|---|
| Discrete states, known actions, need an optimal path | Search (A*, minimax, alpha-beta) | Game playing, routing |
| Variables with constraints, any valid assignment | Constraint satisfaction | Timetables, configuration |
| Uncertain causal relationships | Bayesian networks | Diagnosis, risk scoring |
| Expert rules, need explanations | Expert systems, logic inference | Eligibility, compliance checks |
| Black-box objective, large search space | Genetic algorithms, swarm methods | Design optimisation, hyperparameters |
| Physical agent in an uncertain world | SLAM, control, path planning | Mobile robots, drones |
Modern systems frequently combine both traditions. Language-model agents use search over candidate plans, knowledge graphs ground retrieval, and constraint solvers check outputs that a neural model proposes. Learning handles perception and fuzzy pattern matching; classical methods handle structure, guarantees and verification. Knowing both lets you pick the simplest tool that meets the requirement.
Subsections: Search & Planning, Knowledge & Reasoning, Evolutionary Computation, Robotics.