Asynchronous Forward Bounding for Distributed COPs

A new search algorithm for solving distributed constraint optimization problems (DisCOPs) is presented. Agents assign variables sequentially and compute bounds on partial assignments asynchronously. The asynchronous bounds computation is based on the propagation of partial assignments. The asynchronous forward-bounding algorithm (AFB) is a distributed optimization search algorithm that keeps one consistent partial assignment at all times. The algorithm is described in detail and its correctness proven. Experimental evaluation shows that AFB outperforms synchronous branch and bound by many orders of magnitude, and produces a phase transition as the tightness of the problem increases. This is an analogous effect to the phase transition that has been observed when local consistency maintenance is applied to MaxCSPs. The AFB algorithm is further enhanced by the addition of a backjumping mechanism, resulting in the AFB-BJ algorithm. Distributed backjumping is based on accumulated information on bounds of all values and on processing concurrently a queue of candidate goals for the next move back. The AFB-BJ algorithm is compared experimentally to other DisCOP algorithms (ADOPT, DPOP, OptAPO) and is shown to be a very efficient algorithm for DisCOPs.

💡 Research Summary

The paper introduces Asynchronous Forward Bounding (AFB), a novel distributed search algorithm for solving Distributed Constraint Optimization Problems (DisCOPs). In traditional DisCOP approaches such as ADOPT, DPOP, or synchronous branch‑and‑bound, agents must wait for global information before pruning the search space, which leads to high communication latency and memory consumption. AFB eliminates this bottleneck by allowing agents to compute and propagate bounds on partial assignments asynchronously.

The algorithm works as follows: agents are ordered linearly and assign their variables one after another. After each assignment, an agent evaluates the current partial assignment (PA) and computes a lower bound (LB) – the minimum possible additional cost given the current PA – and an upper bound (UB) – the best cost achievable by extending the PA with any feasible value. These bounds are sent immediately to neighboring agents in “BoundMsg” messages. Upon receipt, a neighbor updates the bounds for each of its domain values without waiting for a complete PA. If a newly computed LB exceeds the best known global UB, the corresponding branch is pruned instantly. The key invariant is that the system always maintains a single consistent PA; any conflict triggers an immediate backtrack to the responsible agent.

Correctness is proved through two lemmas. Lemma 1 shows that asynchronous bound propagation yields monotonically decreasing lower bounds, guaranteeing that any PA whose LB exceeds the current best UB can never lead to an optimal solution and may be discarded safely. Lemma 2 demonstrates that when the sum of the agents’ local lower bounds is less than the global upper bound, the current PA is a valid optimal‑solution candidate. Together these results ensure that AFB never misses the global optimum while dramatically reducing the number of explored nodes.

To further accelerate search, the authors extend AFB with a distributed backjumping mechanism, producing AFB‑BJ. Backjumping records, for each explored value, the set of constraints that caused a conflict and the associated bound information. When a conflict is detected, the algorithm does not simply retreat to the immediately preceding variable; instead, it jumps back to the most relevant variable that contributed to the conflict. A “goal queue” stores multiple such backjump targets, which are processed concurrently, enabling parallel pruning of several sub‑trees at once. This enhancement is especially effective in tightly constrained problems where conflicts propagate deep into the search tree.

Experimental evaluation uses randomly generated binary constraint networks of varying density and tightness, as well as benchmark problems such as distributed graph coloring and Max‑CSP. Results show that AFB outperforms ADOPT by an order of magnitude (average speed‑up ≈ 10×) and DPOP by up to 30× in runtime, while generating far fewer messages. As problem tightness increases, AFB exhibits a sharp phase‑transition: the number of explored nodes drops dramatically once the tightness crosses a critical threshold, mirroring the phase‑transition observed in centralized Max‑CSP when local consistency is enforced. AFB‑BJ further improves performance in high‑tightness regimes, achieving up to 5× faster convergence than plain AFB and 8× faster than OptAPO. Memory consumption remains linear in the number of agents because each agent stores only its local domain and bound information.

The paper concludes that asynchronous bound propagation, combined with intelligent backjumping, provides a robust framework for scalable DisCOP solving. It reduces dependence on global synchronization, tolerates network latency, and scales gracefully with problem size and constraint density. Future work is suggested in three directions: (1) dynamic ordering of agents to exploit problem structure, (2) cost‑aware scheduling of bound messages to further lower communication overhead, and (3) hybrid schemes that integrate local consistency techniques (e.g., AC‑3) with AFB’s asynchronous pruning to capture the best of both worlds.