Mixed-Strategy Flex Distribution

• Mixed-Strategy Flex Distribution is a mechanism in bounded-suboptimal MAPF that hierarchically combines delay-based and conflict-based flex allocation to manage path-cost slack.
• It employs a deterministic sequence—first using DFD, then CFD, and if needed reverting to GFD or zero flex—to ensure both local and global suboptimality guarantees are maintained.
• Empirical evaluations show that MFD improves search focus and CT node expansion efficiency on large benchmark instances, though challenges remain in congested environments.

Mixed-Strategy Flex Distribution is a mechanism for bounded-suboptimal Multi-Agent Path Finding (MAPF) that was introduced within Explicit Estimation Conflict-Based Search (EECBS) as a way to control how path-cost slack, called flex, is allocated during low-level replanning. In that setting, the objective is to retain the bounded-suboptimality guarantee while avoiding a failure mode of greedy flex use: generating constraint-tree nodes whose local bounds remain valid but whose global sum of costs becomes too large for efficient high-level expansion. The mechanism therefore combines Conflict-Based Flex Distribution and Delay-Based Flex Distribution in a hierarchical framework, and is explicitly presented as a conservative control policy over EECBS’s path-cost slack rather than as a game-theoretic randomization rule (Chan et al., 22 Jul 2025).

1. MAPF setting and the role of flex

MAPF is defined on an undirected graph $G=(V,E)$ with agents $a_1,\dots,a_k$, each with a start $s_i$ and a target $l_i$. Time is discrete; an agent can move to an adjacent vertex or wait. The objective is the sum of path costs (SOC), where each path cost is the travel time from start to target, and the bounded-suboptimal goal is a solution with SOC at most $w\cdot C^*$, where $w\ge 1$ and $C^*$ is the optimal SOC (Chan et al., 22 Jul 2025).

EECBS maintains a Constraint Tree (CT). For a CT node $N$, the node lower bound is

$LB(N)=\sum_{i\in[k]} lb_i(N),$

and the node SOC is

$C(N)=\sum_{i\in[k]} c_i(N).$

Its low-level search uses focal search with threshold

$a_1,\dots,a_k$0

where $a_1,\dots,a_k$1 is the minimum $a_1,\dots,a_k$2-value in the low-level OPEN list. Because each path is then individually bounded-suboptimal, EECBS obtains

$a_1,\dots,a_k$3

The paper calls $a_1,\dots,a_k$4 the local suboptimality of $a_1,\dots,a_k$5, while global suboptimality is measured by

$a_1,\dots,a_k$6

where $a_1,\dots,a_k$7. EES expands only CT nodes that are globally bounded-suboptimal, namely those satisfying $a_1,\dots,a_k$8 (Chan et al., 22 Jul 2025).

Flex is introduced because enforcing $a_1,\dots,a_k$9 for every individual path is stronger than necessary. A CT node only needs its total cost to remain within $s_i$0 times the sum of lower bounds. The paper therefore interprets unused slack from other agents,

$s_i$1

as flex that can be reassigned to the replanned agent in a child node. Mixed-Strategy Flex Distribution is the most restrictive of the proposed flex-allocation mechanisms, designed to use that slack without degrading the high-level search focus (Chan et al., 22 Jul 2025).

2. Greedy, conflict-based, and delay-based flex allocation

The baseline flex mechanism in the paper is Greedy-Based Flex Distribution (GFD). When replanning agent $s_i$2 in a child node $s_i$3 of $s_i$4, GFD raises the low-level threshold to

$s_i$5

with

$s_i$6

and

$s_i$7

This guarantees local bounded-suboptimality, but it can make $s_i$8 approach $s_i$9 very closely, which in turn can reduce the fraction of generated nodes that are globally bounded-suboptimal (Chan et al., 22 Jul 2025).

Conflict-Based Flex Distribution (CFD) modifies this by distributing flex in proportion to the number of collisions. If $l_i$0 is the number of conflicts involving agent $l_i$1 in $l_i$2, and $l_i$3 is the total number of conflicts, then for $l_i$4,

$l_i$5

Since $l_i$6, CFD never allocates more than the maximum allowed flex. The rationale is that if flex is meant to help focal search reduce collisions, then agents with more entanglement should receive more slack (Chan et al., 22 Jul 2025).

Delay-Based Flex Distribution (DFD) adds a second signal: estimated delay needed to satisfy the accumulated constraints on the replanned agent. It computes

$l_i$7

where $l_i$8 is the estimated delay for constraint $l_i$9. It then sets

$w\cdot C^*$0

or equivalently,

$w\cdot C^*$1

The paper states that “CFD can be viewed as a special case of DFD where $w\cdot C^*$2.” Ordinary vertex and edge constraints are assigned $w\cdot C^*$3, while corridor and target conflicts receive specific delay estimates derived from the imposed constraints (Chan et al., 22 Jul 2025).

3. Hierarchical construction of Mixed-Strategy Flex Distribution

Mixed-Strategy Flex Distribution (MFD) is the hierarchical mechanism built on top of DFD and CFD. Its purpose is to prevent the over-allocation of flex that can make a child node locally bounded-suboptimal yet globally unattractive for expansion. The procedure is fixed and hierarchical rather than learned (Chan et al., 22 Jul 2025).

If $w\cdot C^*$4, MFD uses GFD immediately. Otherwise it first computes $w\cdot C^*$5 using DFD and checks whether the resulting child would still be globally bounded-suboptimal relative to the current global lower bound: $w\cdot C^*$6 If that condition holds, the DFD allocation is used. If it fails, MFD switches to CFD by setting $w\cdot C^*$7 and recomputing $w\cdot C^*$8. The same global-admissibility test is then applied (Chan et al., 22 Jul 2025).

If both DFD and CFD would still violate the global budget, MFD considers the CT node $w\cdot C^*$9 in LISTs with minimum SOLB, i.e. $w\ge 1$0. It then checks

$w\ge 1$1

and

$w\ge 1$2

If both hold, it redefines the maximum flex as

$w\ge 1$3

If these conditions do not hold, MFD distributes zero flex. The low-level threshold remains of the same form as in the earlier mechanisms,

$w\ge 1$4

but $w\ge 1$5 has passed through this sequence of increasingly conservative filters (Chan et al., 22 Jul 2025).

This hierarchy is the defining content of “mixed strategy” in the paper. It does not denote randomization over several actions; instead it denotes staged selection among DFD, CFD, a stricter global-budget recalculation, and zero flex. This suggests a usage closer to conditional strategy selection than to probabilistic mixing.

4. Theoretical guarantees and invariants

The paper proves that EECBS with GFD is complete and bounded-suboptimal. The proof uses the fact that after low-level search,

$w\ge 1$6

so every generated CT node is locally bounded-suboptimal. It then proves the same conclusion for CFD and DFD, because when $w\ge 1$7, both mechanisms satisfy

$w\ge 1$8

and when $w\ge 1$9, both revert to GFD (Chan et al., 22 Jul 2025).

For MFD, the paper states that EECBS remains complete and bounded-suboptimal because MFD “further eliminates the amount of distributed flex if the maximum allowed flex is positive and switches to GFD otherwise.” The invariant is unchanged: MFD never exceeds the original maximum allowed flex, so it cannot violate the local inequality $C^*$0. Its additional global checks only reduce flex, never enlarge it (Chan et al., 22 Jul 2025).

The paper also gives a key inequality for the third-stage recomputation via $C^*$1. If

$C^*$2

then the redefined $C^*$3 guarantees

$C^*$4

That step is central to the stated purpose of MFD: not only preserving local bounded-suboptimality, but also increasing the chance that the child node remains globally bounded-suboptimal and therefore expandable by EES (Chan et al., 22 Jul 2025).

A common misconception is that any additional flex necessarily improves EECBS. The paper argues against that view directly: although GFD preserves the solution-quality guarantee, increasing thresholds may push the SOC beyond $C^*$5, forcing EECBS to switch among different sets of paths instead of resolving collisions on a particular set of paths. MFD is the proposed response to that specific failure mode (Chan et al., 22 Jul 2025).

5. Empirical behavior, search focus, and failure modes

The evaluation compares EECBS, EECBS-GFD, EECBS-CFD, EECBS-DFD, and EECBS-MFD on six large 4-connected grid benchmarks from the MAPF benchmark suite, with a runtime limit of 120 seconds and standard CBS enhancements including Bypassing Conflict, Prioritize Conflict, and Symmetric Reasoning. The paper explicitly states that “EECBS-CFD and EECBS-DFD already outperform the state-of-the-art EECBS and EECBS-GFD, while EECBS-MFD improves the success rates further” (Chan et al., 22 Jul 2025).

The paper measures global suboptimality and also introduces the CT depth-to-expansions ratio as a measure of how much search effort follows a branch rather than switching among branches. On ost003d with $C^*$6, the averaged ratios are reported as follows: for 100 agents, EECBS 0.99, GFD 1.00, MFD 1.00; for 200 agents, EECBS 0.87, GFD 1.00, MFD 1.00; for 300 agents, EECBS 0.50, GFD 0.97, MFD 0.99; for 400 agents, EECBS 0.18, GFD 0.18, MFD 0.55; for 500 agents, EECBS 0.18, GFD 0.13, MFD 0.22. The reported pattern is that GFD helps at low agent counts but degrades sharply in harder instances, whereas MFD preserves the benefit much better (Chan et al., 22 Jul 2025).

A detailed case study on a 400-agent ost003d instance with $C^*$7 and a 15-second limit shows both EECBS and EECBS-GFD timing out, while EECBS-MFD succeeds. The number of generated CT nodes is 1312 for EECBS, 608 for EECBS-GFD, and 350 for EECBS-MFD; the corresponding numbers of expansions are 843, 395, and 305. The same example reports that, relative to GFD, MFD had a much lower percentage of searches using a large fraction of available flex, with “0% of searches with flex usage over 20” (Chan et al., 22 Jul 2025).

The gains are not universal. In congested MAPF instances on maze-32-32-2 with $C^*$8, the paper reports that without the redesigned low-level FA* search, EECBS-GFD performs worst and EECBS-MFD slightly outperforms plain EECBS; with FA* added, all methods improve, but plain EECBS becomes best. The paper therefore concludes that flex distribution for congested MAPF remains an open question. This limitation is integral to the topic: MFD is presented as the strongest among the tested flex mechanisms for large benchmark instances, but not as a universally dominant replacement (Chan et al., 22 Jul 2025).

6. Terminology and relation to broader mixed-strategy research

Outside bounded-suboptimal MAPF, “mixed strategy” has a broader and more formal meaning. In the theory of mixed-strategy evolutionary algorithms, a mixed strategy is a probability distribution over mutation operators, potentially state-dependent: $C^*$9 In that setting, pure strategy is the special case where one operator receives all probability mass, and mixed strategy is analyzed through a Markov-chain model, asymptotic convergence rate, and asymptotic hitting time (He et al., 2011). A related population-level treatment defines a mixed strategy as a probability distribution of choosing a pure strategy from a set of strategies, and proves that complementarity is the sufficient and necessary condition for a mixed strategy to be never worse than one pure strategy and strictly better for some initial populations (He et al., 2013).

The MAPF notion of Mixed-Strategy Flex Distribution is different. It does not assign probabilities over operators or over flex-allocation rules. Instead, it uses a deterministic hierarchy over several mechanisms. A related distinction appears in industrial 5G scheduling, where a “mixed strategy” over flexible TDD slots is described as “not probabilistic but conditional and adaptive” (Kleinberger et al., 21 Mar 2026). This suggests that the MAPF usage belongs to a broader family of conditional strategy-selection mechanisms rather than to the probabilistic mixed strategies studied in evolutionary computation.

A second misconception is therefore terminological: the phrase “mixed strategy” in MFD should not be read as implying randomized action selection. In the paper where the term is introduced, MFD is a conservative, hierarchical control policy over EECBS’s path-cost slack (Chan et al., 22 Jul 2025).

7. Significance within bounded-suboptimal MAPF

Within MAPF, Mixed-Strategy Flex Distribution is best understood as a mechanism for controlling the interaction between low-level bounded-suboptimal replanning and high-level branch selection. GFD showed that flex can be used to relax per-agent thresholds without losing the global $N$0-bound, but it also exposed a search-control problem: locally valid uses of slack may inflate SOC enough to reduce high-level expandability. MFD addresses that by allocating flex only when doing so remains compatible with branch progress as measured against the current global lower bound (Chan et al., 22 Jul 2025).

Its technical significance lies in how it operationalizes that idea. It first uses DFD, then CFD, then a stricter recomputation based on the minimum-SOLB node, and finally zero flex. That design integrates two distinct signals already present in the search state—collision counts and estimated constraint-induced delay—while preserving the same completeness and bounded-suboptimality guarantees as the more permissive schemes. The empirical evidence indicates that this hierarchy improves success rate and search focus on large benchmark instances, even though congested settings remain unresolved (Chan et al., 22 Jul 2025).

More broadly, Mixed-Strategy Flex Distribution marks a shift in the interpretation of flex. Rather than treating flex simply as spare admissible cost budget to be spent greedily, it treats flex as a resource whose distribution must be coordinated with the global search frontier. That interpretation is the paper’s main conceptual contribution.