A Temporal Planning Approach for Intelligent Flood Response

Fazlul Hasan Siddiqui; Md. Monjurul Islam; Sabah Binte Noor

cs.AIAug 20, 2026

A Temporal Planning Approach for Intelligent Flood Response

Authors: Fazlul Hasan Siddiqui, Md. Monjurul Islam, Sabah Binte Noor

Organizations: Department of Computer Science and Engineering, Dhaka University of Engineering & Technology, Gazipur, 1707, Bangladesh · Department of Computer Science and Engineering, Bangladesh Army University of Engineering & Technology, Natore, 6431, Bangladesh

Abstract

Effective response to multiple, simultaneously flooded areas requires coordinating appropriate actions in the correct temporal order, under severe resource constraints. Automated planning provides a foundation for addressing this challenge by generating time-aware schedules, given a formal description of available resources, constraints, and goals. This work presents an intelligent flood-response framework that exploits temporal planning and models the complete operational life cycle of flood response. The framework incorporates priority-driven triage, route accessibility and travel costs, resource allocation, and supply management, while also supporting mid-execution re-planning in response to unexpected environmental changes. The framework is formulated both in the Action Notation Modeling Language (ANML) and the Planning Domain Definition Language (PDDL) 2.1, facilitating compatibility with a wider range of temporal planners. Experimental results establish the feasibility and scalability of the proposed framework, showing that flood response scenarios can be effectively modeled and solved using temporal planning, while providing guidance on planner selection.

Figures & tables

Figure 1: An example scenario with two safe areas ( safeA and safeB ) and two affected zones ( zoneA and zoneB ). zoneA requires rescue operations, medical support, and 100 units of relief goods, while zoneB requires rescue operations and evacuation of 40 people. Vehicle-dependent travel times range from 11 to 38 time units.

Panel of Figure 2 — Figure 2: A temporal plan for the example flood response scenario in Figure 1 . Each bar represents a durative action, where the horizontal placement indicates the start time and the width with the inscribed number indicates the execution duration in time units. Actions aligned vertically execute concurrently.

Panel of Figure 3 — Figure 3: A compact flood-response instance used as a running example. A single safe location $S$ stages a transit vehicle $T$ , a freight vehicle $F$ , a rescue team $R$ , and a medical support team $M$ , together with a stock of relief goods ( food = k300 ). Zone $A$ outranks Zone $B$ via prior . Edge labels give vehicle-specific travel times.

Action	Preconditions		Effects
[Duration]	ann.	predicate(s)	ann.	predicate(s)
BoardTeam $(t,l,v_{T})$	Start	team_at $(t,l)$	Start	$\neg$ team_at $(t,l)$
[team_size $(t)$ ]	Over-all	vehicle_at $(v_{T},l)$ , $\neg$ evacuating $(v_{T})$	End	team_in_vehicle $(t,v_{T})$
DisembarkTeam $(t,v_{T},l)$	Start	team_in_vehicle $(t,v_{T})$	Start	$\neg$ team_in_vehicle $(t,v_{T})$
[team_size $(t)$ ]	Over-all	vehicle_at $(v_{T},l)$	End	team_at $(t,l)$
MoveVehicle $(v,from,to)$	Start	vehicle_at $(v,from)$	Start	$\neg$ vehicle_at $(v,from)$

Table 1: Transport action schemas. Parameter shorthands:

t

denotes a team;

l

from

, and

to

denote locations;

v_{T}

denotes a transit vehicle; and

T

denotes a vehicle.

Action	Preconditions		Effects
	ann.	predicate(s)	ann.	predicate(s)
RescueAffectedPeople $(l_{A},rescuers,pl)$	Start	prior $(l_{A},pl)$ , $\neg$ needs_rescue $(pl)$ , needs_rescue $(l_{A})$	–	–
	Over-all	team_at $(rescuers,l_{A})$	End	$\neg$ needs_rescue $(l_{A})$
ProvideMedicalSupport $(l_{A},medics,pl)$	Start	prior $(l_{A},pl)$ , $\neg$ needs_medical_support $(pl)$ , needs_medical_support $(l_{A})$	–	–
	Over-all	team_at $(medics,l_{A})$	End	$\neg$ needs_medical_support $(l_{A})$

Table 2: Response action schemas. Parameter shorthands:

l_{A}

denotes an affected location;

pl

denotes a higher-priority affected location; and

C

denotes a constant value.

Action	Preconditions		Effects
[Duration]	ann.	predicate(s)	ann.	predicate(s)
Evacuate $(l_{A},v_{T},pf,pt,pl)$	Start	prior $(l_{A},pl)$ , evac_level $(pl,\texttt{p0})$ , evac_level $(l_{A},pf)$ , $\neg$ evacuating $(v_{T})$	$\textit{Start}^{\dagger}$	$\neg$ evac_level $(l_{A},pf)$ , evac_level $(l_{A},pt)$ , evacuating $(v_{T})$
[person_count_diff $(pf,pt)$ ]	Over-all	vehicle_at $(v_{T},l_{A})$ , valid_evac_step $(v_{T},pf,pt)$	–	–
DisembarkEvacuated-People $(v_{T},l_{S})$	Start	evacuating $(v_{T})$	–	–
[C]	Over-all	vehicle_at $(v_{T},l_{S})$	End	$\neg$ evacuating $(v_{T})$
LoadResource $(r,l,v_{F},kf,kt)$	Start	resource_available $(l,r,kf)$ , $\neg$ cargo_loaded $(v_{F},r)$ , $\neg$ cargo_full $(v_{F})$	Start	$\neg$ resource_available $(l,r,kf)$ , resource_available $(l,r,kt)$

Table 3: Logistics action schemas. Parameter shorthands:

l_{A}

denotes an affected location;

v_{T}

denotes a transit vehicle;

pf

denotes the initial evacuation milestone;

pt

denotes the final evacuation milestone;

pl

denotes a higher-priority affected location than

l_{A}

;

l_{S}

denotes a safe location;

r

denotes a resource;

v_{F}

denotes a freight vehicle;

kf

denotes the initial resource milestone;

kt

denotes the final resource milestone; and

C

denotes a constant value.

Panel of Figure 4 — Figure 4: Dependency among the actions. Transport actions are shown in plain white, response actions in light gray, and logistics actions with a diagonal hatch pattern. Solid arrows indicate causal dependencies; however, the required fact may also be available in the initial configuration. Dashed arrows indicate priority constraints.

Objects
Category	Predicates / values
Location	S (SafeLocation); A , B (AffectedLocation)
Vehicle	T (TransitVehicle); F (FreightVehicle)
Team	R (RescueTeam); M (MedicalSupportTeam)
Resource	food
PersonCount	p0, p20, p40, p60

Table 4: Planning task for the instance of Figure 3 , instantiating the objects, initial state (

I_{f}

), and goal (

G_{f}

) of Definition 7 .

Tier	Locations	Vehicles	Teams	Estimated Actions
1	2 – 6	2 – 3	2 – 5	60 – 554
2	7 – 11	3 – 6	5 – 7	632 – 2560
3	12 – 15	6 – 8	7 – 10	2917 – 7190
4	17 – 25	8 – 12	10 – 13	8566 – 30632
5	26 – 32	12 – 17	14 – 17	32570 – 84579

Table 5: Characteristics of instances grouped by complexity tier; each tier includes 10 problem instances, for a total of 50.

Tier	Planner	Solved	Runtime (s)
Tier	Planner	Solved	Min	Median	Max	Mean
1	FAPE (ANML)	10/10	0.421	0.737	1.709	0.806
1	TFD (PDDL)	10/10	0.001	0.003	0.007	0.003
2	FAPE (ANML)	10/10	1.213	1.811	3.096	1.845
2	TFD (PDDL)	10/10	0.005	0.010	0.017	0.010
3	FAPE (ANML)	9/10	6.378	9.242	28.363	11.938

Table 6: Planner coverage and runtime statistics by complexity. Solved refers to instances solved within the time limit of 3,600 seconds.

Panel of Figure 5 — Figure 5: Percentage of instances solved (coverage) within a given runtime. TFD achieves coverage of 98% within 19.583 seconds, while FAPE plateaus at 86%.

Panel of Figure 6 — Figure 6: Log-log scatter plot of runtime for the instances solved by both planners; shapes distinguish complexity tiers. Points above the dashed diagonal indicate cases where FAPE is slower than TFD.

Planner	Plan Length				Makespan
Planner	Min	Mean	Median	Max	Min	Mean	Median	Max
All Instances
FAPE (ANML)	14	49	39	140	170.0	627.4	533.0	1695.0
TFD (PDDL)	16	71	57	222	222.0	668.5	476.4	3223.8
Jointly Solved
FAPE (ANML)	14	49	39	140	170.0	630.6	536.5	1695.0

Table 7: Plan quality comparison over the instances solved by planners. Plan Length denotes the number of plan steps, and Makespan denotes the total temporal span of execution, with smaller values indicating increased parallelism.

Panel of Figure 7 — Figure 7: Plan makespan comparison for the 42 instances solved by both planners. Most points lie above the diagonal, indicating that FAPE generally produces plans with a higher makespan compared to TFD.

Panel of Figure 8 — Figure 8: Log-log scatter plot of the scalability of both planners with increasing problem size, measured by the size of the grounded action space.

Perturbation	Number of Instances	$\Delta M$ (%)		$\mathbf{\Delta\|\pi\|}$ (%)		$\mathbf{r_{\mathrm{rt}}}$
Evacuee Increase	25	15.7	(6.6)	30.1	(24.3)	0.78	(0.66)
Route Blockage	24	3.2	(0.9)	3.1	(3.1)	0.80	(0.73)
Overall	49	9.6	(4.2)	16.9	(7.7)	0.79	(0.71)

Table 8: Aggregate replanning statistics by perturbation type. Values are means with medians in parentheses.

\Delta M

(%) = percentage change in total makespan;

\Delta|\pi|

(%) = percentage change in plan length;

r_{\mathrm{rt}}

= re-plan runtime divided by original runtime.

Panel of Figure 9 — Figure 9: Original planning runtime versus re-planning runtime (log scale) for all instances, ordered by problem index. Overall, 38 of 49 re-plans were completed faster than the original plan.

Panel of Figure 10 — Figure 10: Re-planning runtime against residual plan length $|\pi_{0}^{\geq}|$ at $T$ (log-scale). The two Re-planning costs are driven by how much of the plan remains to be reconstructed, not by what kind of perturbation triggered it.

Panel of Figure 11 — Figure 11: Percentage change in total makespan after replanning, relative to $\pi_{0}$ ’s original makespan. Route blockage perturbations mostly cluster near zero throughout. Evacuee increase perturbations are far more variable. The dashed horizontal line marks zero change.

Disaster Response

Disaster response uses data, models, and planning tools to assess hazards and coordinate evacuation, relief, and recovery.

See all 59 papers

May 27, 2026

DisasterBench: Benchmarking LLM Planning under Typed Tool Interface Constraints

Disasters cause severe societal impacts, demanding rapid coordination of heterogeneous AI tools, from satellite analysis to flood prediction and damage assessment, into coherent multi-step workflows. As LLMs increasingly serve as orchestrators of such pipelines, effective coordination requires more than selecting semantically plausible tools: LLMs must generate executable workflows with correct parameter binding and dependency propagation. We introduce DisasterBench, a benchmark for evaluating structured multi-agent planning over semantically similar but operationally distinct disaster-response tools. To enable step-level failure attribution, we further propose First-Point-of-Failure (FPoF), which localizes the earliest root cause in a predicted workflow, separating primary errors from downstream cascading effects. Our evaluation reveals three findings: planning method effectiveness depends strongly on model capacity; tool mismatch and parameter-binding errors dominate first failures, revealing semantic grounding and execution consistency as distinct bottlenecks; and verbose intermediate reasoning can create instruction clash with structured output requirements, disrupting plan generation. Together, these findings highlight a fundamental gap between semantic reasoning and execution-grounded coordination, underscoring the need for planning frameworks that jointly model semantic intent, execution constraints, and workflow consistency. Code, data, and evaluation resources are available at: https://github.com/TamuChen18/DisasterBench_Open

Zhitong Chen, Kai Yin, Weifeng Zhang +7

Texas A&M University · University of Toronto

Tool-Use Planning AI Agent Benchmarks Disaster Response LLM Agent Planning

Ask PDF

Jan 8, 2026

Precomputing Multi-Agent Path Replanning Using Temporal Flexibility

Executing a multi-agent plan can be challenging when an agent is delayed, because this typically creates conflicts with other agents. So, we need to quickly find a new safe plan. Replanning only the delayed agent often does not yield an efficient plan, and sometimes cannot even yield a feasible one. On the other hand, replanning other agents may lead to a cascade of changes and delays, and it is computationally expensive. We show how to efficiently replan a single delayed agent by tracking and using the temporal flexibility of other agents while avoiding cascading delays. This flexibility is the maximum delay that the agent can take without changing the order with agents other than the initially delayed agent, or further delaying other agents. Our algorithm, FlexSIPP, precomputes all possible plans for the delayed agent and returns the changes to the other agents within the given scenario. We demonstrate our method in a real-world case study of replanning trains in the densely-used Dutch railway network and in the MovingAI MAPF benchmark set. Our experiments show that FlexSIPP provides effective solutions relevant to real-world adjustments, and within a reasonable timeframe.

Issa Hanou, Eric Kemmeren, Devin Wild Thomas +1

Delft University of Technology · University of New Hampshire

Safe Interval Path Planning Multi-Agent Path Finding Dynamic Replanning Robust MAPF Safe Motion Planning

Ask PDF

Aug 10, 2026

STAIR: Effective Incident Response Using an End-to-End Agentic Planning Framework

Incident response planning is critical for restoring compromised software systems. Existing AIOps methods automate fault diagnosis and remediation and adapt response using feedback, but lack a unified decision basis across planning and execution. This limits their ability to maintain incident state and align actions with evolving recovery objectives during long-horizon, multi-stage security incident response. We present STAIR, an end-to-end agentic response framework that organizes planning and execution around a shared incident state. STAIR maintains incident evidence, response actions, and execution feedback in Graph-as-State (GAS), using the latest state to determine the response stage and recovery objectives for stage-specialized planning. The framework retrieves effect-validated historical experience based on the current state and stage to guide action selection. An Execution Harness executes actions, returns feedback to update GAS, and validates action effects for future experience reuse. We construct STAIR-Bench, an executable incident response benchmark comprising 44 Docker-based cyber-range cases. On its held-out test cases, STAIR achieves an overall defense score of 0.95, improving over the strongest baseline by approximately 10.47%. In the independent CAGE4 environment with ongoing attacks and benign activity, STAIR reduces cumulative penalties by approximately 28.8% relative to the strongest baseline.

Hanlin Jiang, Jionghao Huang, Shaofei Li +6

Key Laboratory of High Confidence Software Technologies, Peking University, Ministry of Education · Southeast University · Huawei Technologies Co., Ltd.

Autonomous Cyber Defense LLM Agent Orchestration Cybersecurity LLM Agent Harnesses LLM Agent Planning

Ask PDF

Abstract

Disaster Response

Explore similar work

DisasterBench: Benchmarking LLM Planning under Typed Tool Interface Constraints

Precomputing Multi-Agent Path Replanning Using Temporal Flexibility

STAIR: Effective Incident Response Using an End-to-End Agentic Planning Framework