Subgoal generation is the cognitive or computational process of breaking down a complex goal into a sequence of smaller, more manageable intermediate objectives. This decomposition enables efficient planning, problem-solving, and learning by reducing cognitive load and providing clear milestones toward the final outcome. Subgoal generation appears in diverse fields, including artificial intelligence (e.g., hierarchical planning and reinforcement learning), psychology (e.g., means-end analysis and chunking), and education (e.g., scaffolding complex tasks). The ability to generate effective subgoals is a key skill in both human reasoning and machine decision-making.
1.1 Means-end analysis and problem space theory
Means-end analysis, a central concept in problem space theory (Newell and Simon, 1972), involves recursively comparing the current state to the goal state and identifying operators that reduce the difference. Subgoals arise naturally when a direct difference-reducing operator is not available; the solver must first achieve an intermediate state that enables that operator. For example, in the Tower of Hanoi puzzle, moving a disk to the target peg requires first moving smaller disks out of the way, forming a chain of subgoals. This decomposition mirrors the core of subgoal generation: breaking a large gap into smaller, actionable steps.
1.2 Cognitive load and chunking mechanisms
Cognitive load theory (Sweller, 1988) posits that working memory has limited capacity. Subgoal generation reduces cognitive load by allowing the problem solver to focus on one intermediate milestone at a time rather than the entire goal. Chunking, a related mechanism, groups lower-level actions into higher-level units (chunks), which can themselves become subgoals. In skill acquisition (Anderson, 1982), procedural knowledge is compiled into productions that automatically activate subgoal sequences, freeing attentional resources for monitoring progress.
1.3 Motivation and goal gradient effect
The goal gradient effect (Hull, 1932; Kivetz, Urminsky, and Zheng, 2006) describes the increase in effort and motivation as one approaches a goal. Subgoals capitalize on this phenomenon by creating multiple “proximity points” that sustain motivation throughout a long task. Each completed subgoal provides a sense of progress, reinforcing continued engagement.
1.3.1 Nearness principle and subgoal salience
The nearness principle states that subgoals that are perceptually or temporally closer to the current state are more salient and more likely to be pursued. Salient subgoals guide attention and reduce the need for extensive search. For instance, in a maze, a visible landmark near the current location becomes a natural subgoal, even if it is not on the optimal path. This heuristic simplifies planning but can sometimes lead to suboptimal sequences if the salient subgoal does not align with the overall objective.
1.4 Developmental aspects in children
Children’s ability to generate subgoals develops gradually with age and metacognitive maturation. Preschoolers often rely on trial-and-error or direct imitation, while older children (ages 7–12) begin to decompose tasks hierarchically. Research on planning (e.g., the “diary of actions” task) shows that by middle childhood, children can spontaneously create subgoals, such as gathering required materials before assembling a model. Training in subgoal formulation improves problem-solving efficiency, suggesting that this skill is amenable to instruction.
2.1 Hierarchical planning in AI
2.1.1 Hierarchical task networks (HTNs)
Hierarchical task networks (HTNs) represent planning domains through tasks (goals) and methods that decompose a high-level task into a network of subtasks. An HTN planner recursively selects methods until primitive actions are reached. Subgoals in HTNs are explicitly defined by the decomposition hierarchy. For example, the task “make coffee” might decompose into subgoals “grind beans” and “brew water.” HTN planners have been used in robotics, logistics, and game AI for their efficiency in structured environments.
2.1.2 Option discovery in reinforcement learning
Options (Sutton, Precup, and Singh, 1999) are temporally extended actions that serve as subgoals in reinforcement learning. An option consists of an initiation set, a termination condition, and a policy. Discovering useful options automatically—without human engineering—is a key challenge. Methods such as the “options framework” and “skill discovery” identify states that are frequently visited, bottlenecks, or goal regions, and then build options that lead to those states, effectively generating subgoals from experience.
2.2 Automated subgoal generation methods
2.2.1 Graph-based abstraction (e.g., subgoal graphs)
Graph-based approaches model the state space as a graph and identify subgoals by analyzing connectivity patterns. Subgoal graphs (e.g., Kaelbling and Lozano-Pérez, 2011) sample states and connect them via feasible paths, then use graph partitioning to find articulation points or densely connected clusters. These points become candidate subgoals. The abstraction reduces the original planning problem to a sequence of easier subproblems, enabling faster planning in large or continuous spaces.
2.2.2 Landmark discovery in planning
Landmarks are facts or states that must occur in any plan that achieves the goal. Landmark discovery algorithms (e.g., Hoffmann, Porteous, and Sebastia, 2004) compute necessary intermediate steps from the domain description. These landmarks serve as natural subgoals. For example, in logistics, “package at airport” is a landmark before “package on plane.” Landmark-based planners use ordered landmarks to guide search, often dramatically reducing the search space.
2.3 Subgoal generation in deep reinforcement learning
2.3.1 Intrinsic rewards for reaching subgoals
Deep reinforcement learning agents can learn subgoal policies by receiving intrinsic rewards for attaining pre-defined or discovered intermediate states. Methods such as “curiosity-driven exploration” (Pathak et al., 2017) define subgoals as states where prediction error is high; reaching them yields a reward. Alternatively, “goal-conditioned” architectures (e.g., Hindsight Experience Replay) treat any visited state as a subgoal for earlier episodes, enabling sparse-reward tasks to be solved through systematic subgoal chaining.
2.3.2 Feudal networks and manager-worker architectures
Feudal networks (Dayan and Hinton, 1993; Vezhnevets et al., 2017) introduce a hierarchy where a high-level manager sets subgoals for lower-level workers. The manager operates at a coarser time scale, emitting a subgoal (e.g., a target state vector) every few steps, and the worker is rewarded for reaching that subgoal. This decoupling allows the manager to learn long-horizon strategies while the worker focuses on short-term control. The architecture has been successful in tasks like navigation and robotic manipulation.
3.1 Scaffolding and guided subgoal decomposition
In educational settings, scaffolding provides learners with temporary support for tasks they cannot yet perform independently. Subgoal generation is a key scaffold: instructors break a complex problem (e.g., writing an essay) into subtasks (thesis, outline, paragraphs). As learners gain competence, the scaffolding is gradually removed, encouraging autonomous subgoal formulation. Research shows that explicit subgoal labeling improves transfer to novel problems, especially in programming and mathematics.
3.2 Intelligent tutoring systems
3.2.1 Dynamic subgoal generation based on learner state
Intelligent tutoring systems (e.g., ACT-R based tutors) model the learner’s knowledge and dynamically generate subgoals tailored to their current skill level. If a learner struggles with a step, the system may insert an extra subgoal to practice a prerequisite skill. Conversely, a proficient learner may receive fewer subgoals, promoting chunking and fluency. This adaptive decomposition optimizes learning efficiency and prevents frustration.
3.3 Gamification and progress tracking
Many games and educational tools use subgoals to maintain player engagement. “Quest” systems break overarching narratives into missions, each with its own completion reward. Progress bars, achievement lists, and level completion thresholds function as explicit subgoals that tap into the goal gradient effect. Subgoal generation in gamification is often designed by designers, but some systems allow players to set personal subgoals (e.g., “defeat 10 enemies”) to foster autonomy.
4.1 Strategies for effective subgoal formulation
4.1.1 Means vs. ends decomposition
A fundamental heuristic is to distinguish between means decomposition (breaking actions into steps) and ends decomposition (breaking the goal into intermediate states). Means decomposition focuses on the steps themselves (e.g., “first do A, then B”), while ends decomposition defines intermediate states that must be achieved (e.g., “first get the key, then reach the door”). Research suggests that ends decomposition often yields more flexible plans because subgoals can be reached by multiple means.
4.1.2 Chunking across domains
Expert problem solvers develop domain-specific chunking patterns—recurring subgoal templates they apply across similar problems. For example, a chess master automatically considers “control the center” and “develop pieces” as subgoals in many openings. Transfer of subgoal heuristics between domains (e.g., from geometry to programming) is possible if the underlying structure is recognized, but often requires explicit training in analogical reasoning.
4.2 Common pitfalls: subgoal fixation and over-decomposition
Subgoal fixation occurs when a solver becomes overly attached to a particular intermediate objective, even when a better path bypasses it. This rigidity can lead to inefficient or failed problem solving. Over-decomposition, or breaking a task into too many tiny subgoals, increases overhead from constant monitoring and switching, negating the cognitive benefits. The optimal granularity depends on task complexity and the solver’s expertise—novices benefit from finer decomposition, while experts prefer coarser ones.
4.3 Relationship to metacognitive monitoring
Effective subgoal generation requires metacognitive skills: the ability to monitor one’s own progress and adjust subgoals as needed. Skilled problem solvers periodically evaluate whether a chosen subgoal is still relevant and whether the decomposition should be revised. This reflection prevents perseveration and adapts to unexpected obstacles. Training in metacognitive strategies, such as self-questioning (“What subgoal am I working on? Is it helping?”), improves overall planning performance.