RedLectern topic
Rules · search · planning · expert systems · combinations with learned models.
For: Learners · developers.
After this page you can: Recognize when explicit rules or search suit a problem.
Page reviewed
Part of: Artificial Intelligence › AI foundations
Symbolic AI represents knowledge in explicit forms, such as symbols and rules, and manipulates those representations to reason or act. Expert systems apply encoded domain knowledge to specific problems. Search explores possible paths toward a goal; planning represents actions and their effects so a system can find a sequence that reaches a goal. These approaches are useful to recognize when a problem has stated rules, goals or constraints that can be represented directly.
The important question is what the problem gives you and what it requires. If the domain has clear conditions and decisions that specialists can encode, rules may make the decision process inspectable. If there are many possible action sequences, search can compare candidate paths, with a heuristic helping direct exploration. When perception or pattern recognition is also needed, neuro-symbolic approaches combine neural and symbolic methods. That combination does not remove the need to check whether the representations, rules and search procedure fit the task.
Watch the lectures on rule-based expert systems and search, then compare the methods with a planning example.
Start with: 6.034 Artificial Intelligence: selected lecture videos
Use the search and planning chapters to see how goals, actions and heuristics become problem representations.
Start with: Artificial Intelligence: Foundations of Computational Agents
Symbolic AI uses explicit representations of knowledge and manipulates them through rules or other algorithms. An expert system applies facts and logical rules to a specific domain; its knowledge is programmed rather than learned from data. This makes the rules available for inspection, while limiting the system to the domain and knowledge that have been encoded.
Use this approach when the relevant conditions and decisions can be stated clearly. For example, a troubleshooting aid could apply a specialist-authored rule such as “if the indicator is off and the device has power, check the connection.” The rule is useful only if it represents the actual domain and the system has the facts it needs.
A concise definition describes expert systems as using domain knowledge and logical rules.
Symbolic AI
AI that represents relevant knowledge explicitly with symbols and manipulates those representations using rules or other procedures.
Expert system
A system that applies encoded specialist knowledge, represented as facts and rules, to draw conclusions in a specific domain.
Heuristic
A practical estimate or shortcut used to guide search or choice; it does not by itself guarantee the best answer.
Neuro-symbolic AI
Approaches that combine neural methods with symbolic approaches to computing and reasoning.
Write one if-then rule for a simple troubleshooting aid. State what fact must be known before the rule can fire.
Example: If the device has power and its status light is off, check whether the display cable is connected. The system must know that the device has power and that the light is off.
Search is a way to explore paths through a problem representation, such as a graph of possible states. A heuristic estimates which options look promising and can guide which path is explored next; depending on the method and the heuristic, a search may fail to find a solution or an optimal one. Planning represents goals and actions, including the conditions under which an action is possible and the state it produces. A planner can then search for a sequence of actions that reaches a goal.
Explicit search or planning is a good fit when a problem can be represented with useful states, transitions and goals. Try it on a route or delivery task: state what counts as a location, which moves are allowed, and what destination is the goal. If you cannot specify those elements, the search procedure has no usable problem to solve.
Introduces states, goals, and action preconditions and effects as parts of a planning representation.
Lectures cover depth-first, hill-climbing, beam, optimal, branch-and-bound, and A* search.
For a delivery robot, name a state, one action with a precondition and effect, and a goal. What would a planner search for?
A state could record the robot’s location and whether it carries the parcel. An action could be “deliver” with the precondition that the robot is at the destination and carries the parcel; its effect is that the parcel is delivered. The goal is delivery, and the planner searches for an action sequence that reaches it.
Neuro-symbolic AI combines neural and symbolic approaches. The two parts can contribute different kinds of computation: a learned model can provide a pattern-based representation, while explicit symbols, rules or search can express structured relations and reason over them. The exact division of work varies by system, so the name alone does not tell you how a particular system is built or whether its reasoning is dependable.
When assessing a proposed combination, ask what the learned part predicts, what the symbolic part represents, and how information passes between them. Then check what happens when an input is uncertain or a rule does not cover the case. These questions help distinguish a system that actually combines learning and explicit reasoning from one that merely uses symbolic language to describe its output.
Provides a comparison point for explicit domain knowledge and logical rules.
Sketch a two-part system for sorting photos into categories: which job could a learned component do, and which job could an explicit rule do?
A learned component could identify objects in a photo. An explicit rule could assign a category when identified objects satisfy a stated condition, such as placing an image in “outdoor scene” when it contains both a tree and sky. The sketch should specify how the object predictions reach the rule.
No recent updates are available.
Stuart Russell and Peter Norvig · 2020
A broad reference that lets readers compare search, logical reasoning, knowledge representation, and planning with other AI approaches.
David L. Poole and Alan K. Mackworth · 2023
The complete text is freely available online. It connects reasoning and planning with machine learning and neural networks in a unified agent framework.
Steven M. LaValle · 2006
A focused algorithmic treatment of discrete and logic-based planning, search methods, and plans; the author makes the full book freely available online.
Pascal Hitzler and Md Kamruzzaman Sarker (eds.) · 2022
A direct overview of neuro-symbolic AI, focused on ways to bridge neural networks and symbolic representations.
Peter Norvig · 1992
Work through classic symbolic AI programs in Common Lisp, including search, logic programming, and expert systems; the author shares the book and code under the MIT license.