Application of Multiagent System and Tabu Search for Truck

Dispatching in Open-pit Mines

Gabriel Icarte Ahumada

1,2 a

and Otthein Herzog

3,4,5 b

International Graduate School for Dynamics in Logistics (IGS), University of Bremen, Bremen, Germany

Faculty of Engineering and Architecture, Arturo Prat University - UNAP, Iquique, Chile

University of Bremen, Germany

Jacobs University Bremen, Germany

Tongji University Shanghai, China

Keywords:

Truck Dispatching, Open-pit Mine, Multiagent Systems, Tabu Search, Scheduling, Rescheduling.

Abstract:

An important and complex process in the mining industry is the material handling process. In this process,

trucks must transport materials extracted by shovels to different places at the mine. To enable efﬁcient material

handling processes, the decision on the destination of a truck is crucial. Currently, this process is supported by

an approach based on centralized systems that apply dispatching criteria. A disadvantage of this approach is

not providing a precise dispatching solution because of missing knowledge about potentially changed external

conditions and the dependency on a central node. We previously developed a multiagent system (MAS-TD)

to solve this problem. In the MAS-TD, intelligent agents that represent real-world equipment interact with

each other to generate schedules. In this paper, we evaluate the MAS-TD by comparing it against a Tabu

Search procedure. In the evaluation, simulated scenarios based on actual data from a Chilean open-pit mine

were used. The results show that both MAS-TD and the Tabu Search procedure are suitable methods to solve

the truck dispatching problem in open-pit mines. However, the schedules generated by MAS-TD are more

efﬁcient than the schedules generated by the Tabu Search.

1 INTRODUCTION

In open-pit mines, the material extracted by shovels

must be transported by trucks to different destinations

at the mine. If a shovel is extracting ore, the extracted

material must be transported by a truck and unloaded

into a crusher or onto a stockpile. If the extracted ma-

terial is waste, it must be transported to a waste dump.

This process is called material handling and it is an

important process for these kinds of mines since the

costs involved in this process can represent up to 50%

of the entire operational cost (Alarie and Gamache,

2002). Figure 1 shows the operations that a truck must

perform from a loading point (shovel) to an unloading

point (crusher, stockpile, or waste dump). These op-

erations are performed repeatedly by a truck until the

shift ends.

Open-pit mines are closed systems where the op-

erations performed are affected by a dynamic envi-

https://orcid.org/0000-0002-1997-0053

https://orcid.org/0000-0003-4781-2551

Figure 1: The truck cycle. Adapted from (Icarte et al.,

2020).

ronment. For instance, equipment failures, changes

in weather conditions or to the state of the routes af-

fect the performance and availability of equipment

and generate delays in material handling (Adams and

Bansah, 2016). In this context and, with the process’s

stochasticity, determining a new destination for the

160

Icarte Ahumada, G. and Herzog, O.

Application of Multiagent System and Tabu Search for Truck Dispatching in Open-pit Mines.

DOI: 10.5220/0010391101600170

In Proceedings of the 13th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2021) - Volume 1, pages 160-170

ISBN: 978-989-758-484-8

truck is not easy.

Currently, the material handling process is sup-

ported by centralized systems based on Operation

Research methods, heuristic procedures, or simula-

tion modeling (Icarte and Herzog, 2019). Most of

these systems use a multistage approach (Alarie and

Gamache, 2002), which computes a guideline in the

ﬁrst stage. A later stage uses this guideline and a dis-

patching criterion (e.g., the current production level

of the shovels) to dispatch the trucks in real-time each

time that the latter requires.

Despite using these solutions, material handling

in open-pit mines is not performed efﬁciently. For

instance, trucks may queue up in front of shovels or

crushers while other shovels wait for trucks. This

generates inefﬁciency, high costs, and not achiev-

ing production targets. Researchers describe that the

current systems don’t solve the problem appropri-

ately because they do not provide a precise represen-

tation of the activities performed by the equipment

(Patterson et al., 2017) and the use of estimated in-

formation (Chang et al., 2015; Costa et al., 2005;

Krzyzanowska, 2007; Newman et al., 2010).

Icarte et al. (Icarte et al., 2020) developed an al-

ternative solution that allows to organize more efﬁ-

ciently the equipment items operations. The solution

is based on a Multiagent system in which intelligent

agents represent real-world equipment. These agents

interact with each other to generate their schedules.

In this paper, we extend the work presented in

(Icarte et al., 2020) by comparing the MAS-TD

against a Tabu Search procedure. Many papers

have demonstrated the application of Tabu Search for

scheduling problems, even for truck dispatching, e.g.,

(Liao et al., 2014). The comparison was performed by

the simulation of scenarios based on actual data from

a Chilean open-pit mine.

The remainder of this paper is structured as fol-

lows: Section 2 presents related work. Section 3 gives

an overview of the developed MAS-TD. Then, Sec-

tion 4 presents the Tabu Search procedure. Section 5

provides the comparison between the MAS-TD and

the Tabu Search procedure and a discussion. Finally,

conclusions and outlook are presented in Section 6.

2 RELATED WORK

The truck dispatching problem in open-pit mines has

been faced in many publications during the last years.

These articles show different methods that try to

achieve two goals: improve productivity and reduce

operating costs (Alarie and Gamache, 2002). For ex-

ample, Ozdemir and Kumral (Ozdemir and Kumral,

2018) developed an agent-based Petri net simulation

model to check whether production targets are feasi-

ble and the extent to control the head grade in min-

eral processing by considering the uncertainties in the

mining operation. Xu et al. (Xu et al., 2019) devel-

oped an Approximate Dynamic Programming (ADP)

algorithm based on Q-Learning. The algorithm im-

plements two models: a static model and a dynamic

model. The static model allocates trucks to a loading

and an unloading job. The model employs an event-

driven method to deﬁne the stage when an idling truck

appears at any time (maybe more than one truck).

Bakhtavar and Mahmoudi (Bakhtavar and Mahmoudi,

2020) developed a two-phase scenario-based robust

optimization (SBRO) model by considering the max-

imization of production, control of ore grade sent to

the crusher, minimization of waiting times for trucks

and shovels, and trucks with different capacities.

Most of the reports on the truck dispatching prob-

lem in open-pit mines follow an allocation model, in

which the destination of a truck is determined when

it is required. Only a few publications have modeled

the problem as a scheduling problem, e.g., Chang et

al. (Chang et al., 2015) and Patterson et al. (Patterson

et al., 2017) proposed algorithms that generate an ini-

tial schedule, which is improved using a metaheuris-

tic method. Their results show that the algorithms

generate schedules for different size instances with

good results and performance in practical time frames

(for the mining industry). Icarte et al. (Icarte et al.,

2020) developed a multiagent system with agents rep-

resenting trucks and shovels. The agents interact with

each other to generate schedules. Their results show

that the MAS-TD provides schedules in practical time

frames and can handle environment dynamics.

3 MULTIAGENT SYSTEM FOR

TRUCK DISPATCHING IN

OPEN-PIT MINES

In the developed MAS-TD, the agents interact with

each other to generate schedules and to maximize the

production at minimum cost. Moreover, the agents

update the generated schedules when a major unfore-

seen event occurs at the mine. Table 1 shows the im-

plemented agents, their objectives, and properties.

3.1 Agent Interactions

In order to generate the schedules, the agents nego-

tiate among them by applying the improved Contract

Net protocol presented in (Icarte et al., 2020). In brief,

Application of Multiagent System and Tabu Search for Truck Dispatching in Open-pit Mines

161

Table 1: Agent description. Adapted from (Icarte et al., 2020).

Agent Real-world

representation

Objective Properties

truckAgent Trucks Create a schedule of the

activities of the truck at

minimum cost

Capacity, loaded velocity,

empty velocity, spotting

time and unloading time,

layout of the mine

shovelAgent Shovels, front

loaders

Create a schedule of the

activities of the equipment

that it represents

considering its target in the

production plan

Capacity, dig velocity, load

velocity and the destination

of extracted material

unloadingPointAgent Crusher,

stockpiles,

waste dumps

Create a schedule of the

activities of the equipment

that it represents

Number of trucks unloading

simultaneously

the shovelAgents start negotiation processes sending

call-for-proposal (CFP) messages to truckAgents. In

the CFP the point of time is offered when the shovel

will be available for loading a truck. TruckAgents re-

spond with proposal messages pointing out the esti-

mated arrival time at the shovel and the cost of per-

forming all the operations involved. Finally, the shov-

elAgent selects the best proposal. The improved Con-

tract Net protocol enables the agents to manage con-

current negotiations. Figure 2 shows this Contract Net

protocol, and Table 2 shows a schedule example for a

truck generated by the agents by applying this proto-

col.

To manage the dynamics of the environment, the

agents follow a complete rescheduling strategy. This

means that the agents generate new schedules from

the point in time where an unforeseen event occur.

The agents cancel their assignments, and they interact

with each other to generate new schedules.

3.2 Decision Making

ShovelAgents must decide what is the best received

proposal. To make this decision, the shovelAgents use

the utility function proposed in (Icarte et al., 2020),

which promotes those proposals that decrease the

shovel’s waiting time and minimize the cost to per-

form the truck operations.

TruckAgents make two decisions. The ﬁrst one is

deciding whether or not to send a proposal to a shove-

lAgent that sent a call-for-proposal message. To make

this decision, the truckAgent checks its schedule and

determines if there is a free time slot for the offered

time in the CFP. If yes, it calculates the total time to

perform all the operations and determines if it ﬁts into

its schedule. If it does not ﬁt into the schedule, it

sends a reject message. If it ﬁts into the schedule,

it sends a proposal.

Figure 2: The interaction between the agents using the im-

proved Contract Net protocol with the conﬁrmation stage.

Adapted from (Icarte et al., 2020).

The second one is to decide on whether or not to

conﬁrm a previously sent proposal. To make this de-

cision, the truckAgent considers the shovel idle time

(informed by the requestConﬁrmation message sent

by the shovelAgent) and the negotiations which the

truckAgent is taking part in. If the shovel idle time is

higher or equal than one minute, the truckAgent sends

an acceptConﬁrmation message. If the shovel idle

time is less than one minute, the truckAgent checks if

it is taking part in another negotiation with more po-

tential beneﬁt for the truck (i.e., with a lower cost to

perform the operations). If yes, the truckAgent sends

a refuseConﬁrmation message. Otherwise, it sends an

acceptConﬁrmation message.

ICAART 2021 - 13th International Conference on Agents and Artiﬁcial Intelligence

162

Table 2: Example of schedule created for a truck.

Assignment Destination Start

Time of

the Trip

Arrival

Time

Start

Time of

the

Spotting

Start Time

of the

Loading or

Unloading

End Time of

the

Assignment

0 Shovel.01 00:32:11 00:55:29 01:02:43 01:03:56 01:06:21

1 WasteDump.02 01:06:21 01:25:23 01:25:23 01:26:03 01:27:23

2 Shovel.04 02:11:24 02:20:27 02:30:35 02:31:10 02:34:29

3 WasteDump.03 02:34:29 02:56:18 02:56:18 02:58:28 03:00:25

4 Shovel.04 03:27:28 03:31:37 03:38:37 03:39:59 03:43:25

5 WasteDump.03 03:43:25 03:59:28 03:59:28 04:01:09 04:02:01

4 TABU SEARCH FOR TRUCK

DISPATCHING IN OPEN-PIT

MINES

Metaheuristics, particularly the Tabu Search (Glover,

1989), are commonly employed to overcome the

complexity of scheduling formulations, especially for

problems dealing with heterogeneous vehicle ﬂeets

(Koc¸ et al., 2016). The Tabu Search algorithm ex-

plores neighborhoods to ﬁnd improved solutions start-

ing from an initial solution. It uses knowledge (tabu

list) about recently visited solutions. The algorithm

rejects neighborhood moves that result in a solution

already in the tabu list to avoid cycling around local

optima. During the algorithm’s main loop, the best

neighborhood that is not in the tabu list is selected

and added to the tabu list. If it is the best solution

found so far, it is recorded as such. The algorithm

terminates after reaching a stop condition and returns

the best solution found.

Algorithm 1 implements the Tabu Search de-

scribed above. In lines 1-5, the initial setup of

the algorithm is done. In sBest the initial solution

provided by Disp-ALG (detail in subsection 4.1)

is stored, and is set as the current solution in best-

Candidate. Also, the tabu list is created and stored

in tabuList. The main loop starts in line 6. This

loop searches for a better solution until a time-out is

reached. In line 7, the function getNeighbors (best-

Candidate) generates new feasible solutions through

small changes in the stored solution in bestCandidate.

Changes in a solution are explained in subsection

4.2. Lines 8 - 13 look for the best candidate among

the neighbors that is not in the tabu list. If this

best candidate is better than the best solution found

so far, it is set as the best solution (lines 14-16).

The best candidate is added to the tabu list (line

17), and if the tabu list is full, the oldest element is

removed (line 18-20). Finally, when the stopTime is

reached, the algorithm returns the best solution found.

Algorithm 1: Tabu Search.

Input: executionTime

Output:

1: sBest ← DISP-ALG

2: bestCandidate ← sBest

3: tabuList ←

4: stopTime ← currentTime+executionTime

5: tabuList.push(sBest)

6: while currentTime<stopTime do

7: sNeighborhood ← GETNEIGH-

BORS(bestCandidate)

8: bestCandidate ← sNeighbor-

Hood.ﬁrstElement()

9: for sCandidate ∈ sNeighborHood do

10: if not (tabuList.contains(sCandidate))

and (ﬁtness(sCandidate) > ﬁt-

ness(bestCandidate)) then

11: bestCandidate ← sCandidate

12: end if

13: end for

14: if ﬁtness(bestCandidate) > ﬁtness(sBest)

then

15: sBest ← bestCandidate

16: end if

17: tabuList.push(bestCandidate)

18: if tabuList.size > maxTabuSize then

19: tabuList.removeFirst()

20: end if

21: end while

22: return sBest

4.1 Initial Solution

Disp-ALG builds feasible schedules for the equip-

ment involved in the material handling process. These

feasible schedules are used as the initial solution in

the Tabu Search algorithm presented above.

The main algorithm (Algorithm 2) consists of 3

steps: the ﬁrst one looks for a shovel task deﬁned as

(s, j, tll). The second step looks for the best truck t to

perform the shovel task. If the second step is suc-

Application of Multiagent System and Tabu Search for Truck Dispatching in Open-pit Mines

163

cessful, the algorithm proceeds with the third step,

which consists of adding the activities and times to

the schedules of the truck t and shovel s; otherwise,

the shovel task is discarded. These steps are repeated

until the last activity of the shovels exceeds the end of

the shift or the target in a production plan is achieved.

To determine whether these conditions are met, the

algorithm invokes the Boolean function isCondition-

sMeet. More details of Disp-ALG can be found in

(Icarte et al., 2020).

Algorithm 2: Disp-ALG.

Input: S, T, Plan, H

Output: shovelSchds, truckSchds

1: shovelSchds ←

0;truckSchds ←

2: while not ISCONDITIONSMEET do

3: s, j,tll ← FINDJOB(S, shovelSchds, H)

4: if s not null then

5: T, act, act

←

FINDTRUCK(T, s, j, tll, truckSchds)

6: if t not null then

7: SCHEDULE(t,s,j,tll,shovelSchds,

truckSchds, act, act’)

8: end if

9: end if

10: end while

4.2 Neighborhoods

The solution space is explored by performing three

operators to generate neighborhoods from the initial

solution. Figure 3 shows examples of generated pos-

sible neighborhoods. The swap operator, shown in

Figure 3(a), generates the ﬁrst one. The swap opera-

tor swaps two truck assignments, ignoring unfeasible

swaps. A feasible swap is only considered if the swap

decreases the cost to perform the assignments. The

second neighborhood, shown in Figure 3(b), is gener-

ated by the change operator. This operator moves an

assignment from one truck to another one only if the

second truck can perform the assignment in a shorter

time than the ﬁrst truck. The new assignment opera-

tor generates the last neighborhood, shown in Figure

3(c). The new assignment operator adds a new assign-

ment by seeking a free time slot to perform a loading

operation in the shovel. Then it looks for a truck to

perform the loading operation.

To select the best neighborhood, an evaluation is

performed based on the efﬁciency of the schedules.

The ratio TotalMaterialToBeTransported / TotalCosts

determines the efﬁciency of schedules. The neighbor-

hood with the highest ratio is selected as the current

solution for an iteration in the Tabu Search algorithm.

Figure 3: Operators to generate neighborhoods.

5 EVALUATION AND

DISCUSSION

The evaluation aims to determine which method is

more appropriate to solve the truck dispatching prob-

lem in open-pit mines. To achieve this, the evaluation

compares the results obtained by applying the MAS-

TD and the Tabu Search procedure on simulated sce-

narios based on actual data.

The evaluation was split into two parts: the ﬁrst

one considers the methods’ performance to generate

schedules. The second one considers how the meth-

ods react to the dynamics of the environment. The fol-

lowing output performance measures are compared in

the evaluation:

• Production: It is the total material transported by

the trucks during a shift. It is measured in tons.

• Costs: Because the actual data does not provide

information on costs such as operator salaries,

fuel costs, and maintenance costs, the truck travel

times are considered costs. They are measured in

hours.

• Required time to generate schedules: It is the

computation time that a method requires to gen-

erate schedules. It is measured in minutes.

• Hauling cost: The ratio between all the material

transported and all the costs to transport these ma-

terials. It is measured in tons per hour.

ICAART 2021 - 13th International Conference on Agents and Artiﬁcial Intelligence

164

5.1 Experimental Setup

In the evaluation, four simulated scenarios are used.

These scenarios are based on actual data from an

open-pit copper mine in Chile. The scenarios con-

sider different heterogeneous ﬂeets of trucks and

shovels in twelve-hour shifts. Table 3 shows some

information about the simulated scenarios.

The actual data, such as velocities and capacities,

is used to set the properties of the agents. Table 4

shows the property sets of the agents. The simulations

ran in PlaSMA (Warden et al., 2010), which is an

agent-based event-driven simulation platform created

for the simulation and evaluation of multiagent sys-

tems. It focuses on simulating logistics processes and

is based on the FIPA-compliant Java Agent Develop-

ment Framework (JADE) (Bellifemine et al., 2007).

All simulations ran on a laptop computer with an In-

tel Xeon 3 Gigahertz CPU, 32 gigabytes of RAM, and

Windows 10.

5.2 Schedule Generation

The evaluation of the schedule generation was split

in two parts. The ﬁrst one aimed to generate sched-

ules for maximizing the production. The second one

aimed to generate schedules considering a hypothet-

ical production plan. In both evaluations, the Tabu

Search procedure ran with different execution times

(5, 10, 15, and 20 minutes).

5.2.1 Schedule Generation for Maximizing

Production

Regarding computation time to generate the sched-

ules, both methods generated schedules in practical

frame times (for the mining industry). Figure 4 shows

the computation time of the methods to generate the

schedules for each shift. MAS-TD generated the

schedules in 20.48 minutes on average. The Tabu

Search procedure generated the schedules in 5, 10, 15,

and 20 minutes as it was pre-set. Also, Disp-ALG was

included in the comparison since it generates the ini-

tial schedules used by Tabu Search. Disp-ALG was

the fastest method to generate schedules. On aver-

age, Disp-ALG required 0,1 minutes to generate the

schedules.

Figure 5 shows the quantity of the material trans-

ported in the schedules generated by the methods.

Regarding production, the schedules generated by

Disp-ALG move the lowest quantity of material in

each of the four scenarios. After 20 minutes run-

ning time, Tabu Search improves the schedules gen-

erated by Disp-ALG by 0.76% on average. MAS-TD

Figure 4: Computation time to generate schedules of MAS-

TD, Tabu Search and Disp-ALG.

is the method whose schedules could move the high-

est amount of material. The activities in its schedules

transported 9.93% more material than the activities in

the schedules generated by Tabu Search.

Figure 5: Production of MAS-TD, Tabu Search and Disp-

ALG.

Figure 6 shows the costs of the generated sched-

ules by the methods for each shift. Regarding costs,

Tabu Search decreases the costs of the schedules gen-

erated by Disp-ALG by 0.92% on average. The

schedules generated by MAS-TD are more expensive

than the activities in the schedules generated by Tabu

Search by 4.59% on average. This is because the

MAS-TD generated schedules with more production,

and therefore more operations to be performed than

the schedules obtained by Tabu Search.

Figure 7 shows the efﬁciency of the generated

schedules by the methods for each shift. Regarding

the efﬁciency of the schedules, determined by the ra-

tio of total Material Transported / Costs, it is a little

bit higher in MAS-TD compared to the Tabu Search

and Disp-ALG, except for the shift 1 and 3, in which

the difference is higher. The generated schedules by

MAS-TD are more efﬁcient than the ones obtained by

Tabu Search by 5.26% on average. Tabu search in-

Application of Multiagent System and Tabu Search for Truck Dispatching in Open-pit Mines

165

Table 3: Simulated scenarios.

Scenario

Number

Trucks

Number

Shovels

Material

Transported

(tons)

Cost

(hours)

Brief Description

1 97 10 455.696,23 780,74 Highest production

2 102 12 378.069,92 772,84 Highest number of

trucks

3 95 13 325.899,61 757,62 Highest number of

shovels

4 96 12 394.759,09 791,59 Many events (truck

and shovel failures)

Table 4: Property values for the simulations.

Equipment Property Unit Min Value Max Value

Trucks

Velocity loaded [km/hr] 20 25

Velocity empty [km/hr] 40 55

Capacity [tons] 230 370

Spotting time [sec] 20 80

Current load [tons] 0 370

Shovel

Capacity [tons] 35 80

Load time [sec] 8 30

Dig time [sec] 8 20

Destination Location at mine (crusher,

stockpile or waste dump)

Crusher Equipment discharging [number of trucks] 1 1

Stockpile Equipment discharging [number of trucks] 1 20

Waste Dumps Equipment discharging [number of trucks] 1 20

Figure 6: Costs of MAS-TD, Tabu Search and Disp-ALG.

creases the efﬁciency of the generated schedules by

Disp-ALG through the improvement operations by

0.93% on average.

5.2.2 Schedule Generation with a Production

Plan

Differently from the previous evaluation, all methods

were set to achieve the targets of a production plan

at minimum costs. The production plan is based on

actual data. Figure 8 shows the computation time

of the methods to generate schedules for each shift.

Figure 7: Efﬁciency of the generated schedules by the meth-

ods for each shift.

Regarding computation times, the methods have the

same pattern as the previous evaluation (for maximiz-

ing the production). It means that Disp-ALG is the

fastest method to generate schedules, MAS-TD gen-

erates schedules in practical time frames (for the min-

ing industry), and Tabu Search generates the sched-

ules in 5, 10, 15, and 20 minutes as it was pre-set.

Compared to the previous evaluation, MAS-TD and

Disp-ALG generate the schedules in less time. This

is because the methods are set to achieve a lower pro-

duction level than in the previous evaluation.

ICAART 2021 - 13th International Conference on Agents and Artiﬁcial Intelligence

166

Figure 8: Computation time of the methods to generate

schedules (with production plan) for each shift.

Figure 9 shows the targets in the production plan,

and the production level reached for the methods for

each shift. Regarding production, all the methods

generate schedules that reach the production level in

the production plan. Due to the fact that generated

schedules by Disp-ALG reach the production level

pointed out in the production plan, Tabu Search does

not improve these schedules (from the perspective of

the production level).

Figure 9: Production level planed and production level

reached by the methods for each shift.

Figure 10 shows the costs of the generated sched-

ules by the methods for each shift. Regarding costs,

Tabu Search decreases the costs of the schedules gen-

erated by Disp-ALG by 1.44% on average. The costs

in the schedules generated by MAS-TD are lower than

the costs in the obtained schedules by Tabu Search by

1.32% on average.

Figure 11 shows the efﬁciency of the generated

schedules by the methods for each shift. Regarding

the schedules’ efﬁciency, it is a little bit higher in

MAS-TD compared to the efﬁciency of Tabu Search

and Disp-ALG. On average, the generated schedules

by MAS-TD are more efﬁcient that the ones obtained

by Tabu Search by 1.39%. Tabu search increases the

Figure 10: Costs of MAS-TD, Tabu Search and Disp-ALG.

efﬁciency of the generated schedules by Disp-ALG

through the improvement operations by 1.46% on av-

erage. This increase in efﬁciency occurs due to the

reduction of the costs calculated by the Tabu Search.

Figure 11: Efﬁciency of generated schedules by MAS-TD,

Tabu Search and Disp-ALG.

5.3 Rescheduling

To evaluate rescheduling, only scenario four was

simulated, because in this scenario many equipment

items experience failures at different times during the

shift. In each event, the methods regenerate the sched-

ules from the moment when the event occurs. The

simulations evaluate the capacity of MAS-TD, Tabu

Search, and Disp-ALG to react when major events oc-

cur at the mine. Tabu search was set to run for ﬁve

minutes to get a quick solution. In this evaluation, the

production reached by the updated schedules and the

calculation time to update the schedules were com-

pared.

This evaluation compared two approaches. The

ﬁrst approach is compound only by the MAS-TD,

which generates the schedules before the shift starts

and generates new schedules each time when an event

Application of Multiagent System and Tabu Search for Truck Dispatching in Open-pit Mines

167

occurs. The second approach is the compound of

Tabu Search and Disp-ALG. In this approach, before

the start of the shift, the schedules are generated by

Tabu Search. Then, for rescheduling, Disp-ALG is

used. Tabu search was not considered for reschedul-

ing because it takes too much time (5 minutes) to react

to the mine’s new conditions.

Regarding computation time to generate the ini-

tial schedules, the ﬁrst approach (MAS-TD) required

around 20 minutes to generate the initial schedules.

The second approach (Disp-ALG and Tabu search)

was much faster and required a little bit more than

ﬁve minutes to generate the initial schedules.

Regarding the reaction to the dynamic of the en-

vironment, the results demonstrate the same previous

pattern: MAS-TD required more time to regenerate

all the schedules in comparison to the DISP-ALG.

Table 5 shows some computation time to update the

schedules at different moment in the shift.

Table 5: Computation time to update the schedules.

Time in

the shift

MAS-TD TS5 Disp-ALG

At the

beginning

0:18:15 0:05:00 0:00:05

In the

middle

0:10:33 0:05:00 0:00:03

At the end 0:03:46 0:05:00 0:00:02

Regarding production, Figure 12 shows the accu-

mulated production during the shift for each method.

The ﬁlled lines represent the planned production

pointed out in the initially generated schedules. Dot-

ted lines represent the reached production in the up-

dated schedules. The graph shows that the methods

can update the schedules when major events occur at

the mine. However, the schedules updated by MAS-

TD achieve signiﬁcantly more production (7.19%)

than Disp-ALG.

6 DISCUSSION

The results demonstrate that MAS-TD, Tabu Search

and Disp-ALG can generate schedules within reason-

able time frames (for the mining industry) whether it

is for maximizing the production or within the target

of a production plan. While Disp-ALG takes only a

few second to generate the schedules, MAS-TD re-

quired around 20 minutes for the scenarios simulated.

The Tabu search computation time depends on the

execution time pre-set as parameter. However, it is

important to mention that the MAS-TD computation

time could be reduced to seconds by employing more

Figure 12: Accumulated production with major events dur-

ing the shift by original and updated schedules.

cores as the MAS-TD architecture lends itself to con-

current computation tasks.

From the perspective of the production tasks, the

schedules generated by the MAS-TD achieve a higher

production rate than the generated schedules by Tabu

Search and Disp-ALG. This is also the case for the

efﬁciency of the schedules. The better results in the

scheduling solution provided by the MAS-TD are due

to the fact that the matching between a shovel task and

a truck is more precise: The Tabu Search starts the im-

provement of the schedules generated by Disp-ALG,

and Disp-ALG follows a sequential creation of the

schedule: it looks for the best truck for a shovel task

and then continues with the following shovel task.

However, this selected truck may be more appropri-

ate for another shovel task. This may cause an inef-

ﬁcient use of the trucks leading to a lower quality of

the schedules. The MAS-TD avoids this situation be-

cause of the concurrent negotiation mechanism. The

mechanism allows for a shovelAgent to decide on the

most appropriate truck’s proposal. Also, a truckAgent

can conﬁrm, or choose to not conﬁrm a previously

sent proposal to a shovelAgent with high idle time.

This allows for a more precise match between shov-

els and trucks.

Considering the results of cost when the methods

generated schedules with a production plan, the costs

of the schedules generated by MAS-TD were a little

bit lower (1.32%) than the costs generated by Tabu

Search. This happens due to the fact that truckAgents

have the chance to select a shovel with a lower cost.

Disp-ALG also looks for a truck to perform the opera-

tions for a shovel in the shortest time. However, when

it ﬁnds it, there is no chance of changing it in the fu-

ture, i.e., if there is another shovel task with less time,

ICAART 2021 - 13th International Conference on Agents and Artiﬁcial Intelligence

168

it is not possible to change it. Tabu search might ﬁnd

this improvement and make the change, but it might

take time to ﬁnd it.

Regarding rescheduling, MAS-TD, Tabu Search,

and Disp-ALG can update schedules when major

events occur at the mine. However, the produc-

tion results and the needed computation time that the

methods take to generate the schedules differs among

them. Disp-ALG is the fastest method to update the

schedules; however, its schedules are the worst in

terms of production rates compared to the generated

schedules by MAS-TD and Tabu Search.

In the simulation, MAS-TD applied a reschedul-

ing strategy for the complete schedule because it aims

at the maximum production level. This strategy takes

much more time than the time required by Disp-ALG

(as is shown in Table 5), because it generates all the

schedules from scratch. However, some assignments

can be made within a short time frame, due to two

reasons: ﬁrst, the algorithms that generate the sched-

ules in MAS-TD are anytime algorithms, i.e., they can

be stopped at any time, and still will generate a fea-

sible schedule solution. Second, the negotiation pro-

cess is fast and many assignments can be computed

within a few seconds. This allows the ﬂeet to con-

tinue working based on a preliminary generated as-

signment, while MAS-TD continues to generate the

schedules. Figure 13 shows that most of the success-

ful negotiations, i.e., assignments computed, are done

in less than one second in all scenarios.

Figure 13: Box and whiskers diagram shows that the mean

of the duration of the successful negotiations is around 0.5

seconds.

7 CONCLUSIONS

A major process in open-pit mines is material han-

dling. In this process, dispatching a truck becomes a

complex decision because of the stochasticity of the

process and the dynamics of the environment. In gen-

eral, a centralized system supports this process. As

an alternative to currently used systems, we present

a multiagent system (MAS-TD) with agents that rep-

resent real world equipment items. In addition, the

agents can react to the dynamics of the mine envi-

ronment and generate new schedules when equipment

items encounter failures.

To evaluate the proposed MAS, we compare it

against a Tabu Search procedure. Our results demon-

strate that the MAS-TD is a more appropriate method

than the Tabu Search procedure to solve the truck

dispatching problem in open-pit mines, because the

MAS-TD generates more efﬁcient schedules than the

Tabu Search procedure. In addition, when critical

events occur in the mine environment, the MAS-TD

regenerates its schedules and achieves a higher pro-

duction than the schedules generated by Tabu Search

and regenerated by Disp-ALG.

In our further research, we will consider more sce-

narios that include other events such as shovel and

truck delays. To decrease the time needed by MAS-

TD for rescheduling, we will evaluate a partial regen-

eration of the schedules.

REFERENCES

Adams, K. K. and Bansah, K. K. (2016). Review of Op-

erational Delays in Shovel-Truck System of Surface

Mining Operations. In 4th UMaT Biennial Interna-

tional Mining and Mineral Conference, pages 60–65,

Tarkwa, Ghana.

Alarie, S. and Gamache, M. (2002). Overview of Solu-

tion Strategies Used in Truck Dispatching Systems

for Open Pit Mines. International Journal of Surface

Mining, Reclamation and Environment, 16(1):59–76.

Bakhtavar, E. and Mahmoudi, H. (2020). Development of

a scenario-based robust model for the optimal truck-

shovel allocation in open-pit mining. Computers and

Operations Research, 115(March 2020):104539.

Bellifemine, F., Caire, G., and Greenwood, D. (2007). De-

veloping Multi-Agent Systems with JADE. John Wiley

& Sons, Chichester, UK.

Chang, Y., Ren, H., and Wang, S. (2015). Modelling and

optimizing an open-pit truck scheduling problem. Dis-

crete Dynamics in Nature and Society, 2015:745378.

Costa, F. P. D., Souza, M. J. F., and Pinto, L. R. (2005). Um

modelo de programac¸

ao matem

atica para alocac¸

est

atica de caminh

oes visando ao atendimento de

metas de produc¸

ao e qualidade. Rem: Revista Escola

de Minas, 58(1):77–81.

Glover, F. (1989). Tabu Search - Part I. Orsa Journal on

Computing, 1(3):190–206.

Icarte, G. and Herzog, O. (2019). A multi-agent system for

truck dispatching in an open-pit mine. In Lottermoser,

B., editor, Abstracts of the Second International Con-

ference Mines of the Future 13 & 14 June 2019,

Institute of Mineral Resources Engineering, RWTH

Aachen University, Aachen, Germany. Verlag Mainz.

Application of Multiagent System and Tabu Search for Truck Dispatching in Open-pit Mines

169

Icarte, G., Rivero, E., and Herzog, O. (2020). An Agent-

based System for Truck Dispatching in Open-pit

Mines. In Rocha, A., Steels, L., and van den Herik,

J., editors, Proceedings of the 12th International Con-

ference on Agents and Artiﬁcial Intelligence - Volume

1: ICAART, pages 73–81, Valleta, Malta. SciTePress.

Koc¸, C¸ ., Bekta, T., Jabali, O., and Laporte, G. (2016). Thirty

years of heterogeneous vehicle routing. European

Journal of Operational Research, 249(1):1—-21.

Krzyzanowska, J. (2007). The impact of mixed ﬂeet hauling

on mining operations at Venetia mine. Journal of the

Southern African Institute of Mining and Metallurgy,

107(4):215–224.

Liao, T. W., Chang, P. C., Kuo, R. J., and Liao, C. J.

(2014). A comparison of ﬁve hybrid metaheuristic

algorithms for unrelated parallel-machine scheduling

and inbound trucks sequencing in multi-door cross

docking systems. Applied Soft Computing Journal,

21(2014):180–193.

Newman, A. M., Rubio, E., Caro, R., Weintraub, A., and

Eurek, K. (2010). A review of operations research in

mine planning. Interfaces, 40(3):222–245.

Ozdemir, B. and Kumral, M. (2018). Appraising produc-

tion targets through agent-based Petri net simulation

of material handling systems in open pit mines. Simu-

lation Modelling Practice and Theory, 87(2018):138–

154.

Patterson, S. R., Kozan, E., and Hyland, P. (2017). Energy

efﬁcient scheduling of open-pit coal mine trucks. Eu-

ropean Journal of Operational Research, 262(2):759–

770.

Warden, T., Porzel, R., Gehrke, J. D., Herzog, O., Langer,

H., and Malaka, R. (2010). Towards Ontology-

Based Multiagent Simulations : Plasma Approach. In

Bargiela, A., Azam Ali, S., Crowley, D., and Kerck-

hoffs, E. J., editors, Proceedings of the 24th Euro-

pean Conference on Modelling and Simulation, ECMS

2010, pages 50–56, Kuala Lumpur, Malaysia.

Xu, T., Shi, F., and Liu, W. (2019). Research on Open-

pit Mine Vehicle Scheduling Problem with Approxi-

mate Dynamic Programming. In 2019 IEEE Interna-

tional Conference on Industrial Cyber Physical Sys-

tems (ICPS), pages 571–577, Taipei, Taiwan. IEEE.

ICAART 2021 - 13th International Conference on Agents and Artiﬁcial Intelligence

170