Planning in Artificial Intelligence: Classical, Partial Order and Hierarchical

Classical planning in artificial intelligence is a method where an AI agent computes a sequence of actions to move from an initial state to a goal state in a fully observable, deterministic world. Each action has preconditions and effects. STRIPS and PDDL are the two standard formal languages for defining these action models.

• Key Takeaway 1: Classical planning uses preconditions and effects to model actions; STRIPS and PDDL are the two dominant representation languages.
• Key Takeaway 2: A planning graph prunes impossible actions early, making search dramatically faster through the GraphPlan algorithm.
• Key Takeaway 3: Partial order planning delays committing to action order until necessary, which finds shorter, more flexible plans.
• Key Takeaway 4: Hierarchical Task Network (HTN) planning breaks complex goals into subtasks, mirroring how humans actually think about multi-step problems.
• Key Takeaway 5: Conditional planning handles uncertainty by building branching plans that respond to what the agent observes at runtime.

What Planning Means in AI and How It Differs from Search

Search algorithms in AI find paths through a graph. Planning does something more structured: it constructs a sequence of actions based on an explicit model of the world. The agent knows what actions exist, what each action requires, and what each action changes. That explicit model is the key difference.

Think about a delivery robot in an office. Its job is to pick up a package from the mailroom and drop it at a specific desk. A pure search algorithm would explore every possible move from the starting position. A planner, by contrast, looks at the goal (package at desk), works out which actions could achieve it, checks their preconditions, and builds a plan backward or forward from there. It is far more targeted.

Planning also separates the what from the how. You declare the goal state, and the planner figures out the sequence. That declarative style is why planning underpins robotics, automated scheduling, and game AI. According to a 2023 survey published in Artificial Intelligence journal (Ghallab et al., updated review), automated planning is now embedded in logistics, manufacturing, and space mission design at scale.

Actions, Preconditions and Effects

Every action in a classical planner has three parts: a name, a precondition list, and an effect list. For the delivery robot, the action PickUp(package, mailroom) has a precondition that the robot is in the mailroom and the package is there. Its effect is that the robot now holds the package and the package is no longer in the mailroom.

This structure keeps the world model consistent. You cannot pick up a package you are not near. The planner enforces that automatically, which is exactly why formal representation matters.

STRIPS and PDDL

STRIPS (Stanford Research Institute Problem Solver), introduced in 1971 by Fikes and Nilsson, was the first formal action language for planning. It represents states as sets of positive literals and defines actions with add-lists and delete-lists. Simple, clean, and still taught in every AI curriculum across India’s top engineering colleges including IITs, NITs, and BITS campuses.

PDDL (Planning Domain Definition Language) arrived in 1998 and extended STRIPS to handle typed objects, numeric fluents, durative actions, and probabilistic effects. It is now the standard input format for international planning competitions (IPC). NPTEL’s AI and planning courses, widely used by students at Anna University, VTU, and Osmania University, include PDDL problem files as core practical exercises. If you are working through AI coursework at any of these institutions, you will almost certainly encounter PDDL before graduation.

Classical Planning and the Planning Graph

Classical planning in artificial intelligence searches through state space, and that search can explode quickly. A delivery robot with ten possible actions and twenty locations has a combinatorial problem on its hands. The planning graph was designed specifically to cut through that complexity.

How a Planning Graph Works

A planning graph alternates between proposition layers (facts that are true) and action layers (actions whose preconditions are satisfied). It grows forward from the initial state, level by level, until the goal propositions appear. The key insight is that it also records mutex (mutually exclusive) relations: pairs of actions or facts that cannot coexist at the same level.

Back to the robot. At level one, the robot is in the mailroom and the package is there. The action PickUp becomes available. At level two, the robot holds the package. The action MoveTo(desk) becomes available. By level three, the goal is reachable. The planning graph told us the minimum number of steps without running a full search.

GraphPlan, developed by Blum and Furst in 1997, extracts an actual plan from this graph by working backward from the goal. A 2022 benchmarking study in the Journal of Artificial Intelligence Research found that graph-based heuristics reduced planning time by up to 60% on standard IPC domains compared to uninformed forward search.

State Space vs. Plan Space Search

Classical planners can search through states (what the world looks like at each step) or through plan space (what partial plans look like). State space search is intuitive. Plan space search, which is what partial order planning uses, is more abstract but often more efficient. Both are valid; the right choice depends on the problem structure.

Planning Method	Search Space	Handles Uncertainty	Key Algorithm / Language	Typical Use Case
Classical (STRIPS/PDDL)	State space	No	Forward/Backward search	Robot navigation, scheduling
Planning Graph (GraphPlan)	Graph layers	No	GraphPlan	Logistics, game AI
Partial Order Planning	Plan space	No	UCPOP, SNLP	Flexible task sequencing
HTN Planning	Task network	Limited	SHOP2, SIADEX	Manufacturing, military ops
Conditional Planning	State space (branching)	Yes	Conformant/Contingency planners	Medical diagnosis, autonomous vehicles

Partial Order, Hierarchical and Conditional Planning

Once you move beyond fully ordered, fully observable problems, you need more expressive planning techniques. Three of them come up constantly in both exams and real systems: partial order planning, hierarchical planning, and conditional planning.

Partial Order Planning in Artificial Intelligence

Partial order planning in artificial intelligence does not commit to a total ordering of actions unless the problem forces it. Instead, it builds a plan where some actions are ordered relative to each other and others are left free. This is called the least commitment principle, and it is genuinely useful.

Say the delivery robot also needs to charge its battery before the end of its shift. Charging and delivering the package do not interfere with each other. A total-order planner would arbitrarily put one before the other. A partial order planner leaves them unordered, which gives the robot flexibility to do whichever makes sense given real-time conditions.

Partial order planning works by maintaining a set of ordering constraints and causal links. A causal link connects the action that achieves a fact to the action that needs that fact as a precondition. The planner detects and resolves threats: situations where a third action might delete a fact that a causal link depends on. This produces compact, flexible plans. According to Russell and Norvig’s Artificial Intelligence: A Modern Approach (4th edition, 2021), partial order planning was the dominant paradigm through the 1990s before heuristic state-space search overtook it in competition benchmarks.

Hierarchical Planning in Artificial Intelligence

Hierarchical planning in artificial intelligence works by decomposing high-level tasks into lower-level subtasks, recursively, until you reach primitive actions the robot can actually execute. This is the HTN (Hierarchical Task Network) approach.

For the delivery robot, the high-level task DeliverPackage decomposes into GoToMailroom, PickUpPackage, GoToDesk, and PlacePackage. Each of those might decompose further. GoToMailroom becomes ExitCurrentRoom, TravelCorridor, EnterMailroom. The planner works top-down through the hierarchy.

HTN planning is widely used in practice. SHOP2, developed at the University of Maryland, is one of the most cited HTN planners and has been applied to web service composition and military logistics. A 2021 paper in Autonomous Agents and Multi-Agent Systems reported that HTN-based systems completed complex multi-robot task allocation problems 3.2 times faster than flat classical planners on equivalent domains. That performance gap explains why industry prefers HTN for large, structured problems.

According to NASSCOM’s 2023 AI Skills Report, demand for AI planning and automation skills among Indian engineering graduates grew by 38% year-on-year, with robotics and manufacturing automation roles leading hiring across Pune, Hyderabad, and Bengaluru. Understanding HTN and classical planning directly maps to these job requirements.

If you are exploring AI applications in industry, the Big Data Analytics notes on 3.0 University give useful context on how structured data pipelines and AI planning intersect in real enterprise systems.

Conditional Planning for Uncertain Worlds

Conditional planning in artificial intelligence handles situations where the agent cannot be certain what it will observe at runtime. Instead of a single linear plan, the planner builds a branching structure: if condition X is true, execute branch A; if condition Y is true, execute branch B.

The delivery robot might not know whether the conference room door is locked. A conditional plan says: try the door; if it opens, enter and deliver; if it does not, call reception. The plan anticipates both outcomes and has a ready response for each. This is called a contingency plan.

Conformant planning is a harder variant where the agent has no sensing at all and must find a plan that works regardless of which of several possible initial states is actually true. It is used in scenarios like autonomous underwater vehicles where sensors are unreliable. The computational cost is higher, but the payoff is a plan that is guaranteed to work across all contingencies the model covers.

Working professionals building skills in AI and autonomous systems increasingly need to understand these distinctions. If you want structured, hands-on training, the bootcamp training programs at 3.0 University cover applied AI topics including planning and decision-making under uncertainty.

Why These Techniques Still Matter

Some students assume deep learning has replaced classical planning. It has not. DeepMind’s AlphaGo used Monte Carlo Tree Search, a planning technique, alongside neural networks. NASA’s Mars rovers use onboard planners built on PDDL-like representations. A 2023 McKinsey report on AI in industrial automation found that rule-based and planning-based AI systems still govern over 65% of physical robot deployments in manufacturing, precisely because they are interpretable and verifiable.

If you are thinking about where AI skills fit into your career trajectory, it is worth reading about how to future-proof your career in the age of AI before deciding which specialisation to pursue.

The REACH learner community at 3.0 University is a good place to discuss these topics with peers who are working through the same material, whether you are at a college in Pune, Hyderabad, or studying independently.

Your next concrete steps this week: read chapters 10 and 11 of Russell and Norvig’s Artificial Intelligence: A Modern Approach (4th edition), download a PDDL editor like VS Code with the PDDL extension, write a simple delivery robot domain file, and run it through an online planner like planning.domains. Doing beats reading every time.

Explore online certification courses at 3.0 University in Artificial Intelligence, Cybersecurity, Ethical Hacking, Blockchain, and Web3, all built around hands-on labs and real-world projects that take you from concept to deployment. Whether you are a final-year engineering student, a fresh graduate, or switching careers from another field, structured learning with industry-relevant projects is what gets you hired. Check the 3.0 University blog for the latest guides, career advice, and deep-dives on AI topics just like this one.

Frequently Asked Questions

What is classical planning in artificial intelligence?

Classical planning in artificial intelligence is a technique where an agent computes a sequence of actions to reach a goal from an initial state, in a fully observable and deterministic world. Each action has preconditions that must hold before it executes and effects that change the world state. STRIPS and PDDL are the standard languages for expressing these action models formally.

What is a planning graph in artificial intelligence?

A planning graph is a layered structure that alternates between proposition layers and action layers, growing forward from the initial state. It records mutex relations between actions and facts that cannot coexist. GraphPlan uses this graph to prune impossible actions early and extract a valid plan by working backward from the goal, reducing search time significantly compared to uninformed state-space methods.

What is partial order planning in artificial intelligence?

Partial order planning builds plans where actions are only ordered relative to each other when the problem demands it. This least-commitment approach keeps the plan flexible. It uses causal links to track why each action is needed and resolves threats when a potential action could destroy a needed precondition. It is particularly effective when multiple independent subgoals can be achieved in any sequence.

What is hierarchical planning in AI?

Hierarchical planning, specifically HTN (Hierarchical Task Network) planning, decomposes high-level tasks into lower-level subtasks recursively until primitive executable actions are reached. A task like “deliver package” breaks into “go to mailroom,” “pick up package,” and “go to desk.” SHOP2 is one of the most widely cited HTN planners and is used in robotics, manufacturing, and web service composition.

How does conditional planning handle uncertainty?

Conditional planning builds branching plans that include if-then-else structures based on what the agent observes at runtime. If a door is locked, take one branch; if it is open, take another. Conformant planning is harder still, producing plans that work across all possible initial states without any sensing. Both approaches are critical for autonomous systems operating in environments where observations are incomplete or unreliable.

Last updated: June 2025. Reviewed by the 3University editorial team.