An agent that pursues several goals has to keep them consistent with what it believes. If it learns the bridge is out, it should stop trying to cross it. If it later learns the bridge was repaired, it should take the goal up again. The best formal account of this, Shapiro and Brewka's situation-calculus agent, ties the set of active goals directly to the current beliefs. This has a blunt consequence. The only way to bring a goal back is to change a belief.
Real deliberation is often subtler. The agent may still believe the bridge is out but recognise that this fact does not block a goal served by a drone. It may keep a sensor's alarming reading on record but judge that this particular reading should not cancel that particular goal, because the sensor is flagged as unreliable for that kind of measurement. What gets disputed in these cases is not the belief. It is the link from the belief to the goal.
Argumentation with recursive attacks has exactly this kind of object. In these frameworks an "attack" is a directed conflict between two arguments, and an attack can itself be attacked. This note proposes an agent in which goals are arguments and "belief $b$ makes goal $g$ impossible" is an attack. Arguments that a particular impossibility claim does not apply here are attacks on that attack. The agent's active goal set is defined as the goals in the unique, most cautious acceptable set of the resulting graph, called the grounded extension. Suspension and readoption then fall out of defeat and reinstatement, without retracting beliefs.
The two-sentence version is as follows. Today, goal–belief consistency treats "this belief rules out that goal" as a fixed fact, so goals can only return when beliefs change. We make that ruling-out a defeasible, goal-specific object that can be undercut, and we test whether this gives more correct and more explainable goal management when beliefs are noisy or apply only in some contexts.
Core Claim
Our claim is that goal-specific undercutting of belief–goal conflicts should be the unit of revision in a rational agent. Doing so preserves the consistency guarantees of the classical account and fixes a concrete failure mode of it.
The failure mode arises whenever one belief is relevant to some goals but not others, or is trusted for some purposes but not others. A belief-only revision scheme must either retract the belief, which is wrong for the goals it does block, or duplicate it into goal-indexed variants, which grows the knowledge base and hides the reason.
We make two technical commitments. First, we use AFRA's grounded semantics, which is unique and computable by iterating a monotone function to a fixpoint. The agent's chosen goal set is therefore deterministic, like Shapiro and Brewka's. It also never requires enumerating many extensions, the operation that broke argumentation solvers on planning-derived graphs. Second, we prove a conservativity result: with no undercutting attacks present, and with goal priorities encoded as attacks on goal-conflict attacks, the grounded goal set coincides with the output of Shapiro and Brewka's makecons filter under a total priority order. If that correspondence fails, the proposal is in trouble at the foundations, which is why we check it first.
Where This Sits
The proposal takes its machinery from two sources. The first is AFRA: Argumentation framework with recursive attacks. AFRA lets attacks target attacks, extends Dung's semantics to sets of arguments and attacks, guarantees a unique grounded extension, and maps back to ordinary flat frameworks. The second is Modgil's Labellings and Games for Extended Argumentation Frameworks. That paper encodes "A is preferred to B" as an attack on B's attack on A, and we reuse the idea to express goal priorities.
The target behaviour comes from Dynamic Interactions between Goals and Beliefs. In that work, chosen goals are recomputed from beliefs by the priority-ordered makecons filter, with proved consistency (Theorem 5.1), belief–goal compatibility (Theorem 5.2), and automatic readoption when beliefs change. We keep these properties. What we change is that readoption can also be triggered by defeating the conflict rather than revising the belief, which that formalism cannot express.
The closest prior combinations motivate the gap without filling it. Decision Support through Argumentation-Based Practical Reasoning already uses AFRA attacks-on-attacks so that a value preference breaks a symmetric conflict between treatments. It evaluates a static, single-step decision, however, and has no belief update or goal readoption loop. Normative Practical Reasoning via Argumentation and Dialogue and Normative Practical Reasoning: An Argumentation-Based Approach compile plans, goals and norms into flat frameworks. They resolve overrides with a counting tie-breaker (the "Democratic Principle") or with grounded acceptance over whole plans, and they leave mid-execution revision open. Defeasible Normative Reasoning: A Proof-Theoretic Integration of Logical Argumentation shows that grounded acceptance in a flat framework matches a well-understood normative consequence relation. This is encouraging for our choice of grounded semantics, but that work models no goals.
Two results constrain the design. On Structural Properties of Argumentation Frameworks: Lessons from ICCMA reports that about 23% of complete-semantics enumeration instances in ICCMA 2017 were solved by no solver within 600 seconds, with planning-derived graphs failing because of layered cyclic structure. This is why we avoid enumeration. Higher-Order Argumentation Frameworks: Principles and Gradual Semantics shows that recursive formalisms diverge from one another once acceptance is numerical. This is why we restrict the core claim to grounded AFRA and treat graded strengths, for example for actions moving through the "Functional / Suspect / Deprecated" self-model of Adaptable and Verifiable BDI Reasoning, only as a secondary arm.
Research Program
Graph built at each step. At each step the agent builds an argument graph from four ingredients:
- Goal arguments, one per requested goal.
- Belief arguments, from its current belief state.
- Impossibility attacks. A belief argument attacks a goal argument when, under the agent's action model, no believed course of action satisfies the goal given that belief. This is the argumentative counterpart of B-consistency.
- Conflict attacks, symmetric attacks between jointly unsatisfiable goals.
Higher-order attacks. Two kinds of attacks act on these attacks. Priority attacks target a lower-priority goal's attack on a higher-priority one, following Modgil. Undercutters target a specific impossibility attack. An undercutter is generated from three sources: applicability rules ("this impossibility concerns ground transport; this goal's plans use air"), the agent's sensor or action self-model in the style of Stringer et al. ("the source of this belief is Suspect for this goal's precondition"), and explicit normative permissions.
Goal selection and update. The active goal set is the set of goal arguments in the AFRA grounded extension. When a belief, rule or self-model state changes, the agent updates the graph and recomputes the grounded extension incrementally from the affected region.
Theoretical results. We will prove that active goals are always jointly consistent. We will prove that an impossibility attack that is not undercut always suspends its goal, which is the analogue of Theorem 5.2. We will prove the conservativity correspondence with makecons stated above. We will also characterise readoption: a goal is reinstated exactly when every surviving attack on it is defeated, either directly or by undercutting.
Explanations. The explanation for any suspension or readoption is the chain of defeats in the grounded labelling. We render it as a proponent–sceptic dialogue of the kind Shams et al. defined for flat frameworks.
Baselines
The naive reference point is a blindly committed agent. It adopts all requested goals by priority once and never drops one unless it is achieved, a caricature of rigid commitment from Intention = Choice + Commitment. It establishes how much revision matters in each benchmark at all.
The strongest existing method for the question we ask is Shapiro and Brewka's makecons filter, implemented over the same action model and fed exactly the same belief stream. It is the strongest because it is the only cited method with proved belief–goal consistency and automatic readoption. We give it every advantage in two variants. The first revises beliefs by simple overwrite. The second places a reliability-weighted belief filter in front of makecons, and that filter receives the same sensor-fault and self-model signals that generate our undercutters. Both variants are given the same per-step compute limit and the same hand-tuning budget for thresholds. Beating it means higher per-step active-goal accuracy and fewer wrongful suspensions at no larger knowledge-base size. If the filtered variant matches us only by encoding goal-indexed duplicate beliefs, we will count the knowledge-base size difference as the result and report it as such.
The third baseline is utility-optimal replanning with goal values minus norm penalties in answer-set programming (ASP), as in Practical reasoning with norms for autonomous software agents. We rerun it at every belief change. Its goal values and penalties are tuned on held-out episodes with the same budget we use for our applicability rules.
The fourth baseline is a flat-framework agent in the style of Shams et al. that uses the Democratic Principle in place of priority and undercutting attacks.
Two ablations isolate the key ingredient. The first removes undercutters and converts each one into a retraction of the corresponding belief, so that it becomes makecons expressed in argumentation form. The second keeps our graph but solves it after AFRA's mapping to a flat framework with a mature solver. This second ablation should reproduce our decisions exactly and differ only in runtime. That is the point of running it.
For none of these baselines does the research file contain a reported number on a dynamic goal-revision task. We found no reported result on Lacuna for makecons, for ASP normative replanning, or for flat-framework normative agents on any quantitative benchmark, and the disaster-recovery, clopidogrel and Socratic-dialogue scenarios are reported only qualitatively. The one quantitative anchor relevant to computation is the roughly 23% universal-failure rate on ICCMA 2017 complete enumeration within 600 seconds reported by Rodrigues et al.
Why not just…?
Why not just make the belief filter smarter and keep makecons? This is the real threat. When every belief is relevant to every goal in the same way, a reliability filter in front of makecons should match us, and we expect it to. The difference appears only when trust or applicability is goal-specific: one Suspect sensor reading should still block a goal whose plans depend on it critically, but should not block a goal with a redundant route. A belief filter must then either pick one global verdict or duplicate the belief per goal. Our benchmark varies the fraction of goal-specific undercutting from zero upward, so that the advantage, if it exists, appears as a slope rather than an anecdote. If the slope is flat, the representation is not earning its keep and we will say so.
Why not just flatten AFRA and use ordinary argumentation solvers? For grounded semantics we think this is a fine backend, and Baroni et al. prove that the mapping preserves the semantics. We therefore do not claim new decision power from recursion at the solver level. The flattening ablation exists to confirm identical decisions. The claim is about the agent's representation: impossibility relations are first-class, goal-specific and revisable, and that is what produces the encoding-size and explanation differences. Only under the secondary graded arm do flat reductions break the principles studied by Amgoud et al., and there we use their m-avg semantics directly.
Why not just optimise a utility function in ASP? Shams et al. show that this computes utility-optimal plans soundly. It can, however, only express "this goal no longer applies" by retuning numbers, and it gives no defeat chain explaining a readoption. We measure both the accuracy gap and explanation length. If the tuned utility agent matches our accuracy, the remaining argument is explainability alone, and we would report the proposal's contribution as narrower.
Proposed Experiments
Benchmark. Because the file contains no benchmark for dynamic goal revision, we build one. Its generators lift the disaster-recovery domain of Shams et al. and the clopidogrel scenario of Cerutti into multi-step episodes, plus a parameterised synthetic domain. Each episode has a hidden true world state. From that state an oracle independent of any argument graph labels which goals are achievable at every step. The belief stream injects three kinds of event: genuine impossibilities, faulty-sensor reports, and context-limited facts that block only some goals. We control the rates of each.
Metrics. We measure per-step active-goal accuracy against the oracle, wrongful suspensions, missed suspensions, goal churn (spurious drop-then-readopt cycles), knowledge-base size needed to reach a given accuracy, update latency, and length of the defeat chain behind each change.
Decisive checks. The first check is the conservativity test. On episodes with no undercutting events, our agent and makecons must select identical goals at every step. Any disagreement is a bug or a falsified theorem.
The second check is the slope test. As the fraction of goal-specific undercutting grows, our advantage over the filtered makecons should grow while holding knowledge-base size fixed. A flat curve refutes the core claim.
The third check is scalability. We generate layered, cyclic graphs patterned on the Planning2AF structure Rodrigues et al. identified, add recursive attacks, and compare grounded fixpoint computation against complete-semantics enumeration under the ICCMA 600-second limit. We expect grounded computation to stay polynomial where enumeration hits the failure regime they report.
Secondary arm. In the secondary arm, actions degrade through the Functional, Suspect and Deprecated states of Stringer et al. and undercutter strengths vary. Here we test whether HO-AF m-avg strengths rank goals for readoption better than the binary grounded labelling, using the same oracle.
References
- AFRA: Argumentation framework with recursive attacks
- Labellings and Games for Extended Argumentation Frameworks
- Dynamic Interactions between Goals and Beliefs
- Decision Support through Argumentation-Based Practical Reasoning
- Normative Practical Reasoning via Argumentation and Dialogue
- Normative Practical Reasoning: An Argumentation-Based Approach
- Defeasible Normative Reasoning: A Proof-Theoretic Integration of Logical Argumentation
- On Structural Properties of Argumentation Frameworks: Lessons from ICCMA
- Higher-Order Argumentation Frameworks: Principles and Gradual Semantics
- Adaptable and Verifiable BDI Reasoning
- Intention = Choice + Commitment
- Practical reasoning with norms for autonomous software agents