Team-Subset Partial Replanning in Multi-Agent Systems

• Team-subset partial replanning is a strategy that modifies only a subset of agents' plans to reduce computation and communication costs in complex systems.
• Utility partitioning, partial preference models, and decentralized algorithms enable stable, resource-efficient adaptations under uncertainty.
• Algorithms like CBBA-PR, D-MAPF with ASP, and reactive LTL replanning showcase practical applications in robotics, UAVs, and reinforcement learning.

Team-subset partial replanning refers to the class of strategies, models, and algorithms that modify or re-plan only a subset of agents’ plans or policies within a larger team, rather than requiring the entire team to globally recompute their actions or allocations. This principle arises in multi-agent systems, cooperative game theory, multi-robot coordination, planning under uncertainty, and organizational dynamics, reflecting practical realities where full-team replanning is prohibitively complex, both computationally and communicationally. Research has produced several normative and algorithmic frameworks for identifying, enabling, and theoretically analyzing such subset-focused replanning, drawing on utility partitioning, formal logic, decentralized consensus, constraint programming, and learning theory.

1. Subset Utility and Stability in Cooperative Team Games

The foundation for team-subset partial replanning is formalized in the subset team games framework, which generalizes the classic transferable utility (TU) model to allow each agent subset $A$ to evaluate the outcome of any coalition $S$ according to its own utility function $u_A(S) = u_A(V(S))$, where $V(S)$ is the outcome produced by $S$ (0907.2376). This structure enables the decomposition of overall marginal contribution as

$m_A(A \cup B) = u_{A\cup B}(A\cup B) - u_B(B)$

for disjoint $A, B$. The self-interested (competitive) and altruistic components are separated as:

• Competitive: $$c_A(A \cup B) = u_{A\cup B}(A \cup B) - u_B(A \cup B)$$,
• Altruistic: $a_A(A \cup B) = u_B(A \cup B) - u_B(B)$, such that $m_A = c_A + a_A$.

This decomposition maps subset contributions within a "cooperation space," identifying subteams that, under partial replanning (only $A$ or $B$ adjust), will remain stable and aligned if both $c_A$ and $a_A$ are positive (Quadrant I). Instability and potential for subteam defection, requiring selective replanning or incentive design, occur when altruistic contributions are negative. This analytic partition directly informs which subteams are natural candidates for partial replanning interventions.

2. Planning and Replanning with Partial Preferences

Partial preference models motivate and technically underpin team-subset partial replanning in centralized and distributed planning scenarios (Nguyen et al., 2011). When user (or agent subset) preferences are incompletely specified—e.g., as a convex combination $$V(p, w) = w\cdot\mathrm{time}(p) + (1-w)\cdot\mathrm{cost}(p)$$ parameterized by $w$—optimal plans for each possible preference realization define a spectrum of candidate solutions. Planners then seek plan sets that cover the space of plausible user or subteam interests, as measured by quality metrics such as: $$\mathrm{ICP}(P^*) = \sum_{i=1}^k \int_{w_{i-1}}^{w_i} h(w) \left(w\cdot\mathrm{time}(p_i) + (1-w)\cdot\mathrm{cost}(p_i)\right) dw$$ where $P^*$ is the set of plans optimal for some trade-off $w$.

This supports subset-oriented replanning: rather than requiring the global team to agree on a single compromise plan, representative sets are constructed whose constituent plans satisfy the dominant preferences of different subteams, facilitating partial replanning and distributed choice among interdependent agents.

3. Algorithmic Mechanisms for Subset-Scoped Replanning

Specific algorithms have been advanced to operationalize team-subset partial replanning:

• CBBA-PR: In decentralized task allocation, the Consensus-Based Bundle Algorithm with Partial Replanning (CBBA-PR) (Buckman et al., 2018) allows each agent to "reset" only the tail (lowest-value tasks) of its assigned bundle when reallocating upon the arrival of new tasks, instead of a full reset. Furthermore, only a subset of the team may participate in the partial replan, reducing communication burden and convergence time. Intermediate strategies, such as partial team or local resets, provide trade-offs between speed and overall coordination quality.
• D-MAPF with ASP: In dynamic multi-agent path finding, new agents or environmental changes are handled by identifying minimal conflict sets; Answer Set Programming (ASP) solvers selectively replan paths for only those agents involved in collisions, keeping unaffected plans fixed (Atiq et al., 2020). The ASP framework enforces hard constraints (no collision within the replanning subset) and soft weak constraints (minimal interference with the fixed set), enabling efficient minimal-disruption conflict resolution.
• Reactive LTL Replanning: Planning under temporal logic constraints with heterogeneous teams involves decomposing a global task into independently assigned subteams via SMT methods (Leahy et al., 2020) or, in case of local robot failures, reassigning only the failed subtasks among functioning team members, with theoretical guarantees of minimal violation with respect to user-specified subtask priorities (Kalluraya et al., 22 Oct 2024). The graph search for failed subtask reassignment ensures only the necessary subteam replans, and plan segments are updated locally if sufficient overlap (reused execution) is present, else global resynthesis occurs.

4. Decision and Coordination via Dynamic Subsets

Team-subset partial replanning is conceptually analogous to subset-based collective decision-making in swarms (SubCDM) (Fuady et al., 1 Aug 2025). Here, consensus on environmental features is reached by a dynamically constructed, locally determined decision-making subset $S_{\mathrm{DM}}$, rather than by the entire swarm. Local recruitment (via leader-based hop counts or distributed confidence adaptation) adjusts the subset's size depending on the convergence difficulty, while the rest of the swarm remains available for other parallel tasks. This strategy generalizes to partial replanning by localizing adaptation and resource expenditure to the agents directly involved in new information processing or changed circumstances.

5. Distributed, Reactive, and Learning-Based Partial Replanning

Distributed replanning algorithms for multi-robot teams (RLSS (Şenbaşlar et al., 2021), RecBayes (Ribeiro et al., 18 Jun 2025)) and hybrid planning frameworks integrate local sensing, partial observability, and local plan validity checks to allow each robot or subteam to iteratively replan their trajectories or policies without global coordination. RLSS utilizes linear spatial separation constraints over Bèzier control points to safely enable per-robot trajectory repair in response to dynamic obstacles or neighboring path changes, with no inter-robot communication. RecBayes employs a recurrent Bayesian classifier to recognize team-task structure from partial observations, enabling agents to adapt (replan their own contribution) solely based on locally available data and without full awareness of the team state.

In multi-agent reinforcement learning, sub-team policy curriculum and domain-of-expertise (DoE) classification (Fosong et al., 2023) allow only the non-expert or affected subset to be unfrozen and retrained, while other agents maintain their previously learned behaviors. Hyperparameters (exploration, regularization) are selectively modulated based on expertise assessments, efficiently focusing adaptation on necessary subteams and maintaining overall joint policy stability.

6. Implications, Tradeoffs, and Theoretical Guarantees

Team-subset partial replanning frameworks offer key advantages:

• Resource Efficiency: Limiting replanning (and communication or computation) to subsets of a team leads to substantial savings in decentralized settings (Buckman et al., 2018, Fuady et al., 1 Aug 2025).
• Scalability: Decomposition via utility function, logic specification, or learning curriculum enables parallel and independent subteam updates, supporting large heterogeneous teams (Leahy et al., 2020, Fosong et al., 2023).
• Dynamic Robustness: Reactive, subset-scoped replanning ensures system resilience to failures or unexpected changes without full-team disruption (Kalluraya et al., 22 Oct 2024, Zhou et al., 2021).
• Theoretical Guarantees: Soundness, completeness, local/global minimality, and violation optimality are provided in various models through explicit formulas (e.g., minimum-violation costs in LTL task reallocation (Kalluraya et al., 22 Oct 2024)) and graph-search/optimization arguments.

Tradeoffs naturally arise between the degree of global optimality (achievable by full replanning) and the communication, computational, or operational cost of larger-scale recoordination. Frameworks that enforce strong independence or local-only updates occasionally reduce overall robustness or flexibility in reallocations (Leahy et al., 2020).

7. Application Domains and Future Directions

Team-subset partial replanning methodologies are increasingly prominent in:

• Dynamic multi-UAV or multi-robot task allocation under online information and resource constraints,
• Multi-agent path finding in dynamic and adversarial environments,
• Team restructuring in organizational or social network settings with volatile membership (Li et al., 2021),
• Swarm robotics requiring adaptive consensus without global participation,
• Large-scale reinforcement learning where task decomposition or partial policy retuning is critical.

Ongoing research explores tighter integration of partial preference modeling, decentralized logic planning, reactive learning, and real-time theoretical guarantees to accommodate increased system heterogeneity, stronger environment coupling, and online adaptation to emergent scenarios. A general theme is the identification and use of minimal, functionally critical subteams to efficiently and reliably propagate necessary changes throughout complex cooperative structures.