IPOCL Planner: Intent-based Narrative Planning

• IPOCL planner is a narrative planning algorithm that integrates explicit character goals and frames of commitment to generate plausible, intention-driven story actions.
• It extends traditional partial-order causal link planning by embedding motivation constructs, ensuring all non-happening actions are explained through character intentions.
• Empirical evaluations using the QUEST model demonstrate IPOCL’s improved reader comprehension and consistent inference of true character motivations.

The Intent-based Partial Order Causal Link (IPOCL) planner is a refinement search narrative planning algorithm designed to generate action sequences that are not only causally coherent but also support audience comprehension of character intentionality. IPOCL extends traditional partial-order causal link planning (POCL), integrating reasoning about character goals alongside plot achievement to produce narratives in which every character action is explainable via explicit intention, thereby enhancing believability and understandability of the story structure (Riedl et al., 2014).

1. Formalization of the IPOCL Narrative Planning Problem

IPOCL formalizes the narrative planning problem as a tuple $(I, C, G, \mathcal{A})$:

• $I$ denotes the initial world state, specified as a set of ground propositions (“fluents”).
• $C$ is a finite set of symbols representing narrative characters.
• $G$ is the author’s desired outcome, also a set of propositions.
• $\mathcal{A}$ is a library of STRIPS-style action schemata, each containing:
  • Actors: subset of parameters denoted as intentional agents
  • Happening: a Boolean ($\text{true}$ indicates events that may occur without intention)
  • Constraints: ground literals that must be present in $I$ and are immutable
  • Preconditions: literals required for action execution
  • Effects: literals added/deleted, including special “intends(a, p)” effects

Actions applied to the world state transition $I$ towards a state in which all $p \in G$ are satisfied.

2. Data Structures: Plans and Frames of Commitment

IPOCL plans extend the classical partial-order/causal-link (POCL) representation by introducing explicit character intention constructs. In conventional POCL, a plan is a 4-tuple $(S, B, O, L)$:

• $I$0 is a set of instantiated steps, including “start” and “goal” sentinels.
• $I$1 is a set of binding constraints (variable unifications).
• $I$2 is a set of ordering constraints ($I$3).
• $I$4 is a set of causal links $I$5, meaning $I$6 achieves $I$7 for $I$8.

IPOCL augments this with a set $I$9 of Frames of Commitment (FoC), yielding $C$0. Each FoC is a 5-tuple $C$1, where:

• $C$2 denotes all steps performed by character $C$3 in pursuit of $C$4
• $C$5 is the internal goal, a literal, pursued by $C$6
• $C$7 is the terminal step whose effect is $C$8
• All $C$9 precede $G$0 in $G$1

Every non-happening action must reside within at least one FoC, enforcing that all deliberative character actions are intention-driven.

3. IPOCL Refinement-Search Algorithm

IPOCL employs least-commitment search to incrementally construct and repair plans. The algorithm introduces four flaw types: A. Open preconditions: unsatisfied preconditions, as in standard POCL B. Causal threats: interference between causal links C. Open motivation flaws: FoCs lacking a step establishing “intends(a, g_a)” D. Intent-flaws: steps that may or may not belong to a FoC

High-level pseudocode follows a classic flaw-selection loop, with refinements:

• Causal planning repairs open preconditions by introducing or reusing steps, spawning new FoCs, or associating steps to existing ones, along with corresponding motivation flaws and causal links.
• Motivation planning repairs open motivation flaws by instantiating a step whose effect is “intends(a, g_a)” and enforces temporal ordering.
• Intent planning repairs intent-flaws by assigning steps to a FoC (propagating intent-flaws backward across the character's actions) or discarding the association.

Completeness is achieved when no open preconditions, causal threats, open motivations, or orphaned steps (unexplained intentional actions) persist.

4. Character Intentionality vs. Conventional Partial-Order Planning

Conventional POCL narrative planners treat all actions solely as means to achieve the author-specified $G$2, making no distinction between authorial and character goals. In contrast, IPOCL discriminates between the external “outcome” $G$3 and internally discovered character goals $G$4. Characters acquire intentions through “intends(a, g_a)” effects, spawning FoCs that explain the motivation behind each action.

Every non-happening action taken by a character must be subsumed within at least one FoC, enforcing that all such actions are interpreted as motivated by explicit, plot-internal goals. This construct prevents implausible narrative artifacts that arise when characters execute actions unreasonable for their narrative role solely to achieve the author’s external objective (e.g., a princess assassinating the king just to fulfill “king dead” as $G$5).

5. Empirical Evaluation and Audience Comprehension

IPOCL’s effectiveness in rendering character intentions comprehensible was empirically evaluated using the QUEST question-answering model. Two narratives, both generated from the same world description—one by IPOCL and one by POCL—were converted to natural language and administered in a controlled study.

• Participants were provided multiple “why did X do Y?” queries with corresponding answers; all pairs were presented (52 in POCL, 82 in IPOCL).
• Goodness-of-Answer (GOA) was rated on a 1–4 scale.
• QUEST was used to label pairs as “good” or “poor,” correlating to plan structure.

Results indicated that, for “good” pairs, IPOCL narratives achieved mean GOA of 3.20 (compared to 2.99 for POCL), and for “poor” pairs, IPOCL mean GOA was 1.19 (vs. POCL’s 1.30). IPOCL stories were thus judged more consistently with QUEST predictions, supporting the claim that IPOCL enables readers to reliably infer and endorse true character motivations, while rejecting implausible motives more readily (Riedl et al., 2014).

6. Theoretical Properties and Computational Complexity

IPOCL’s refinement search increases both branching factor and potential search depth compared to POCL planners. Specifically, for flaw resolution:

• Let $G$6 = number of possible groundings of an operator schema
• Let $G$7 = number of effects per operator
• Let $G$8 = number of intentional actors per operator
• The branching factor per step is $G$9, where the “$\mathcal{A}$0” accounts for the option to forgo FoC creation and exponent $\mathcal{A}$1 opens distinct goals for multiple actors

Search depth $\mathcal{A}$2 is governed by the count of all possible flaws (open preconditions, open motivations, intent-flaws, threats), resulting in an overall complexity $\mathcal{A}$3 for some constant $\mathcal{A}$4. Solving even modestly complex stories can require strong heuristics; for example, generating an Aladdin-style narrative required approximately 12 hours and expanded 1.8 million nodes.

---

IPOCL represents a significant extension of narrative planning, embedding explicit character psychology and goal-driven action selection directly into the plan representation and refinement search. The result is narrative plans that not only advance an author’s desired outcome, but that support audience comprehension of plausible, intention-driven character behavior and address limitations of conventional POCL approaches (Riedl et al., 2014).