Reliable, Distributed Scheduling and Rescheduling for Time-Critical, Multiagent Systems

Abstract

This paper addresses two main problems with many heuristic task allocation approaches-solution trapping in local minima and static structure. The existing distributed task allocation algorithm known as performance impact (PI) is used as the vehicle for developing solutions to these problems as it has been shown to outperform the state-of-the-art consensusbased bundle algorithm for time-critical problems with tight deadlines, but is both static and suboptimal with a tendency toward trapping in local minima. This paper describes two additional modules that are easily integrated with PI. The first extends the algorithm to permit dynamic online rescheduling in real time, and the second boosts performance by introducing an additional soft-max action-selection procedure that increases the algorithm's exploratory properties. This paper demonstrates the effectiveness of the dynamic rescheduling module and shows that the average time taken to perform tasks can be reduced by up to 9% when the soft-max module is used. In addition, the solution of some problems that baseline PI cannot handle is enabled by the second module. These developments represent a significant advance in the state of the art for multiagent, time-critical task assignment. Note to Practitioners-This work was motivated by the limitations of current agent-to-task allocation algorithms that do not use a central server for communication. In previously published work, the current state-of-the-art consensus-based bundle algorithm has demonstrated poor performance when applied to model task allocation problems with critical time limits, often failing to assign all of the tasks, especially when the deadlines are tight. The performance impact (PI) algorithm has a much better success rate with these model problems but would be flawed when applied to real missions because it has no mechanism for online replanning when new information becomes available. In addition, it is somewhat restricted in the way it searches for a problem solution, meaning that more efficient plans are often available but are not discovered. This paper tackles both of these shortcomings. The PI algorithm is extended to include a module that permits rescheduling when necessary, and a further module is introduced that widens the scope of the solution search. A third module that is able to offer robust plans, even for large-scaled missions involving many agents and tasks, has also been developed, although it is not discussed here. Implementation

Addressing robustness in time-critical, distributed, task allocation algorithms

“…Two particular combinatorial auctionbased algorithms lend themselves to the solution of the problems of interest in this paper -CBBA [4] and the PI algorithm [3]. It has been shown empirically that the baseline PI algorithm performs better than the baseline CBBA algorithm [3,8,26] for deterministic task allocation problems with tight task deadlines, with PI demonstrating a much better success rate with different numbers of tasks and agents, and different network topologies. However, the papers mentioned do not examine PI's handling of uncertainty.…”

“…The lack of a robust solution is a major shortcoming, as time-critical missions (such as SAR) demand a low probability of failure. Reassignment, as discussed in [8], can deal with some of the problems, but it can be hard to recover from a situation where rescue vehicles have already been sent in the wrong direction.…”

“…The problem of interest is documented fully in [3,8,26] and [38]. It is the optimal, conflict-free assignment of a set of n heterogeneous agents V = [v 1 , .…”

“…The removal performance impact (RPI) measures the benefits of removing a task from a bundle and the inclusion performance impact (IPI) measures the benefits of adding a task. Full details of the metrics and the PI algorithm are presented in [3,8,26,38]. The details are not reproduced here as, for the purposes of this paper, the reader only needs to know that there is a consensus phase, when the agents agree on the schedule, and that the RPI and the IPI are calculated using the time costs c i,k associated with each agent i and task k; the candidate robustness modules are created by using different robust versions of those costs (see Section 3.3).…”

“…It also requires only 100 samples compared with the 10,000 needed in Ponda's robust CBBA model [7], and is tested using higher numbers of tasks and agents. ROB-M can also be bolted on to baseline CBBA, and it is shown that, for uncertain problems, the failure rate and number of unallocated tasks also reduces, as long as the CBBA-ROB-M algorithm is solving a problem that baseline CBBA can solve deterministically; CBBA generally has difficulty when time constraints are tight or the task to vehicle ratio is too high [8]. PI with ROB-M is also compared directly with Ponda's robust CBBA algorithm and demonstrates clear superioriority in terms of mean numbers of solved tasks.…”