Human-agents Interactions in Multi-Agent Systems: A Case Study of

Human-UAVs Team for Forest Fire Lookouts

Sagir Muhammad Yusuf and Chris Baber

∗

University of Birmingham, B15 2TT, U.K.

Keywords:

Human-agent Team, Mixed-initiative Reasoning, Bayesian Learning, Mixed-initiative Planning.

Abstract:

In this paper, we propose an architecture that uses predictions tools obtained via Bayesian learning algorithms

to monitor the issues of communication, fault tolerance, and adaptation in human-agent mission. The archi-

tecture describes different level of knowledge, planning, and commands differ by their priorities. We tested

the model using forest ﬁre lookouts problem on a simulation platform (AMASE). The process uses the conju-

gate gradient descent algorithm to perform the Bayesian Belief Network training. The output of the training

process is a well-trained BBN for agents’ prediction, estimation, and decision making during communication

failure. The prediction perfection of the human and agents were compared and studied. Although results proof

that human approach is prone to error but is good in terms of emergency commands execution. We suggested

that the use of a well-trained prediction tool (i.e., the output BBN) could be used in monitoring mission during

communication link, hardware, or software breakdown.

1 INTRODUCTION

The human-agents team is applicable in many real-

world applications such as health care system (Kifor

et al., 2006), customer service system (Tecuci et al.,

2007), disaster management (Cacace et al., 2014), etc.

During the interaction, both participants need to be

supportive to each other in order to have a collabora-

tive system. A fully collaborative system tries to bal-

ance the knowledge and reasoning between the agents

and the human experts. Unlike intelligent assistant

where human has knowledge but consult agent for

manipulation purposes or its opposite tutorial assis-

tant (Rich and Sidner, 2007). In order to achieve that,

several issues exist such as effective task distribution,

communication, awareness, control, structuring, eval-

uation, adaptation (Makonin et al., 2016; Tecuci et al.,

2007), and fault tolerance. We here discuss the prob-

lems sequentially.

Task distribution refer to the way of segment-

ing the task and choosing the task performer (Turpin

et al., 2014). That is the issue of who does what

task. The problem becomes complicated in hetero-

geneous agents management because of the presence

of different types of agents. Communications refer to

the mode of receiving commands and knowledge be-

∗

https://www.Birmingham.ac.uk/staff/proﬁles/computer-

Science/baber-Chris.aspx

tween the collaborative participants (i.e., human-and

the agents) during the mission (Makonin et al., 2016).

It mostly occurs via the use of Graphical User Inter-

face (GUI) structured in what-next approach. Aware-

ness refers to how the human and agents understand

the current situation. That is the level of knowledge

of the participants about the current condition of the

environment (Tecuci et al., 2007). A more advance

approach is reasoning the current situation and pre-

dicting the near future events known as the Situation

Awareness (Endsley, 1995). Control refers to who

will take over initiative at a particular time (Tecuci

et al., 2007). It is the question of who is the cur-

rent boss. Structuring issues refers to the architec-

tural design of the human-agent system. That is how

the component of the system was organised (Makonin

et al., 2016). Evaluation refers to the critical analy-

sis of the participants’ tasks and the expected beneﬁts

in-between the human and the agents (Tecuci et al.,

2007). Adaptation refers to how the agents and the

human learn their environment and mode of opera-

tions (Makonin et al., 2016). The seven aforemen-

tioned aspects refer to the essential aspects of human-

agent interaction, they relied on an effective commu-

nication platform. We suggest the consideration of

hardware or software failure during the interaction

process that could result in communication failure.

These challenges were itemised and solved in

Yusuf, S. and Baber, C.

Human-agents Interactions in Multi-Agent Systems: A Case Study of Human-UAVs Team for Forest Fire Lookouts.

DOI: 10.5220/0009369203470354

In Proceedings of the 12th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2020) - Volume 1, pages 347-354

ISBN: 978-989-758-395-7; ISSN: 2184-433X

347

many kinds of literature works with the assumptions

of reliable hardware, software, and communication

link(Bevacqua et al., 2015; Cacace et al., 2014; Fer-

guson and Allen, 2007). In this paper, we recommend

the consideration of hardware or software breakdown

(i.e., fault tolerance) that could result in communica-

tion and awareness updates failure. Fault tolerance is

the ability of the team to manage any unexpected inci-

dent such as communication link breaking, hardware,

or software error, and all other sort of Byzantine fail-

ures (Araragi, 2005). Therefore, we intend to answer

the following questions in human-agent interaction.

• Is there any approach for tackling communication

failure in human-agent team?

• In the presence of faulty communication link,

hardware, or software, could agents and human

team maintain the same environmental awareness

through appropriate adaptation?

The ﬁrst question was thrown on the effect of com-

munication failure during the human-agent mission

and how it affects both agent and human knowledge

of the environment. This could be categorised into

two levels. (1) agent-to-agent updates problem, espe-

cially in a multi-agent system, and (2) human-agent

update. For instance, one could wonder the outcome

of human-UAV surveillance mission with communi-

cation failure in the middle of the mission. The mis-

sion could face the following challenges:

• Incomplete data and command management: it

happens when an agent or human proposes a com-

mand, then the command was not successfully re-

ceived or executed due to communication failure

occurrence.

• Agents power management.

• Lack of supportive knowledge: because of com-

munication breakdown, the agents and human will

lack supportive advice. For example, if the agents

are familiar with their environment, they can be

able to suggest some supportive ideas and insights

about any destructive actions of human experts

and vice versa.

To address the aforementioned issues, we propose a

model architecture that utilises collaborative activities

in a human-agent team using Bayesian inference and

learning. Its constraints are the absence of commu-

nication and the prediction accuracy of the predictive

tools for effective system Situation Awareness main-

tenance.

2 RELATED WORK

In human-agent team, agents are capable of making

huge computations, navigation, and large data collec-

tion, etc. within a short period while the human expert

is intelligent enough to extract the information, and

control the mission cognitively (Makonin et al., 2016;

Rich and Sidner, 2007; Tecuci et al., 2007). A chal-

lenge therefore arises in controlling the human-agent

activities such as control, task allocation, communi-

cation, etc within the system.

Different architectures were proposed in manag-

ing aspects of human-agent team management. For

instance, system control can be managed using par-

allel or full execution. In parallel control, the agents

and the human experts learn concurrently and evalu-

ate their outcomes against the objective function satis-

factions on a time-to-time basis (Tecuci et al., 2007).

After the parallel execution, the agents and the hu-

man learn their errors an avoid them in future mis-

sions based on prioritised tasks. Example of sys-

tems that implement such techniques are Diamond-

Help and Collagen (Rich et al., 2001; Rich and Sid-

ner, 2007) .

Overall control involves the precise observation of

the co-participant’s actions and learns from their ac-

tions. The user may guide the agent about the current

situation of the environment, while at the background;

the agent is learning and correcting its errors and mis-

takes. The same thing goes to the human expert in the

absence of enough knowledge; the agent can guide

him/her using a what-next strategy as in RESIN (Yue

et al., 2010), PerCon (Su, ), 2014), ForceSpire (Endert

et al., 2014), and ALIDA (Green et al., 2010). Other

aspects, such as adaptation, awareness, control, evalu-

ation, and system design, were addressed in (Makonin

et al., 2016; Tecuci et al., 2007).

This paper suggests the consideration of system

components failures that could result in communica-

tion, awareness, and adaptation problems. For exam-

ple, imagine a communication failure in human-UAV

team mission at a separation distance of 2 kilometres

with critical battery conditions. How does the agents

that rely on human-expert for control could save the

mission by continuing with the tasks and ensure per-

fect mission delivery? Is there any balanced platform

for monitoring the teams at the absence of the com-

munications links? We pay more emphasis on this

issue and propose a model that will maintain the bal-

ance between the two mission’s participants. This

model uses an accurate predictive tool that handles

uncertainties and runs a parallel system with synchro-

nisation, unlike the traditional approach of recovery.

Regarding awareness handling, DiamondHelp and

HAMT 2020 - Special Session on Human-centric Applications of Multi-agent Technologies

348

Collagen system (Rich et al., 2001; Rich and Sidner,

2007) uses a chat window to maintain the current un-

derstanding of the environment. While the STC sys-

tem learns from the previous users’ experience in line

with commands execution. The propose model varies

by focusing more attention on the accuracy of the pre-

diction tooland current situation awareness analysis

of the system during components failure. The model

tackles the awareness problem using two approaches:

• Analysing current situation.

• Making effective predictions and estimations on

the future situation.

It is similar to the concept of Situation Awareness in

(Endsley, 1995; Stanton et al., 2006) but formalise

with inferential reasoning and fault tolerance. Next

section describes the model in details.

3 THE PROPOSE MODEL

This section introduces the proposed model and ar-

chitecture, which intends to approach the problems of

awareness and fault tolerance in human-agent team.

It uses Bayesian inference and learning in solving the

issues. Figure 1 describes the model architecture.

Figure 1: The Propose Model Architecture.

The proposed model has three entities, that is, human,

agent, and the intermediary server labeled A, B, C in

ﬁgure 1. We here below explain all these components.

3.1 Human Expert (a)

The human-expert refers to a professional person that

has knowledge of controlling, structuring, and execut-

ing the mission. For example in ﬁre ﬁghting, he/she

could be a well-experienced ﬁre warden ofﬁcer that

witness many complex and simple cases. In a case

where agents are exposing to a new environment, the

human expert could have more priority in controlling

the mission’s plans, commands, and data ﬂow.

3.1.1 Human Plans (A1)

These are set of to-do actions by the agents propose

by the human expert. It could be a queue or stack

of actions to be performed during the mission by the

agent. For instance, in ﬁre lookouts missions, it could

be a set of places to be visited in sequential order.

3.1.2 Human Commands (A2)

Current executing plan is refer to as a command. This

module is responsible for monitoring the human com-

mands and prioritising them. It monitors the plans

execution states such as completed, interrupted, non-

executable. A command could be mark as complete

when it was opportune to execute all its actions. An

interrupted command is a command that was removed

during execution due to encountering of higher prior-

ity command or lack of available resources (e.g., bat-

tery and computational power) to execute the com-

mand.

3.1.3 Human Knowledge (A3)

Human knowledge refers to the expert’s suggestions

and actions for controlling the system. It could be

inputted via the Graphical User Interface (GUI) and

pass to the hidden server (C) or agent (B) via com-

munication link. Human or agent knowledge could

be updated or prioritise best on the evaluation by the

team in optimising the objective functions.

3.1.4 Human Metadata (A4)

All updates records about the mission are recorded as

the metadata. This could be used to traced the mis-

sion and also be used for training purposes. It could

be exposed to learning algorithms such as gradient de-

scent, expectation maximisation, counting algorithms

etc. (Bottou, 2010; Dempster et al., 1977; Romany-

cia, 2019) to obtain a well-train BBN for making ac-

curate predictions, estimation and decisions.

3.1.5 Human Prediction Tools (A5)

Human predictive tools refers to the well-trained tools

(e.g, trained BBN) for helping the expert in making

predictions estimations, and conclusions. That is, it is

the output of the training process using the available

knowledge. It could be used to estimate or predict

what the agent is doing during the mission or in near

future.

Human-agents Interactions in Multi-Agent Systems: A Case Study of Human-UAVs Team for Forest Fire Lookouts

349

3.2 The Agent (B)

Agent refers to any autonomous hardware entity that

is capable of helping the expert in achieving the mis-

sion. For example, UAV, wheel-robots, legged robots

etc. As an autonomous entity, agent is capable of gen-

erating its plans and executes them. It could also re-

ceive other plans and commands from the human ex-

pert.

3.2.1 Agent’s Plans (B1)

As an autonomous entity, agents could have some

strategies of generating and monitoring their plans by

following certain algorithms. At times agents and hu-

man plans contradict each other when that happens,

the model proposes C1 (ﬁgure 1) to prioritise and de-

cide on which plan to be executed.

3.2.2 Agent’s Commands (B2)

Agents execute their plans in a certain structures such

as queue, stack, etc. The conﬂict between an agent

and human commands could be resolved by the pri-

oritiser C1.

3.2.3 Agent’s Knowledge (B3)

Agents’ knowledge comes from the sensor data,

which could be organised in a situation-based man-

ner. The agent’s knowledge is manage by the human

expert or the prioritser (C1 ﬁgure 1) by using different

techniques.

3.2.4 Agent’s Metadata (B4)

During mission execution, agents’ sensors informa-

tion and other valuable data about the environment

(e.g., time, location, etc.) are recorded as metadata to

the agents. These data can be used for learning pur-

poses using any suitable learning algorithms.

3.2.5 Agent’s Prediction Tools (B5)

This is a set of predictions tools such as well-trained

Bayesian Belief Network (BBN) and neuro-fuzzy

system to be used by the agent in making predictions,

estimations, and decisions during the absence of com-

munication link between it and the human expert.

3.3 Connector Server (C)

The connector server comprises the computer that

provides the Graphical User Interface (GUI) for com-

munication and the communication link. For exam-

ple, in human-UAV mission, it could be the linking

the agent, human expert, PC or mobile phone (as de-

scribed by ﬁgure 2).

Figure 2: Structure of Human-agent Interaction.

Figure 2 describes the global view of ﬁgure 1, It

shows the human labelled as (A), the agent (B), and

the connector PC (C). The communication link (e.g.,

wireless Local Area Network) connects the human

expert, connector server, and the agent.

3.3.1 Prioritiser (C1)

Prioritiser assigns a probability values to plans and

commands for agents’ execution. It a very important

module residing in the connector server. It is capable

of doing that because of the following reasons:

• It receives the knowledge of both human and the

agent. When human sends a command, the server

monitors the agent’s actions on the command and

learns the whole interaction.

• It has high computational capacity.

• It controls the communication link.

Therefore, by considering the aforementioned rea-

sons, the prioritiser have enough resources for priori-

tising tasks. But the human expert also have an ability

to execute an emergency plans.

3.3.2 Knowledge (C2)

The server has the global view of the systems because

it receives data from both human and agents. It moni-

tors and assesses its knowledge based on the satisfac-

tion of the command’s output. The server knowledge

could simply be referred to as the combination of both

the agents and human knowledge contributions.

3.3.3 Tools and Algorithms (C3)

This module comprises of the software to be used in

structuring knowledge for learning process such as

HAMT 2020 - Special Session on Human-centric Applications of Multi-agent Technologies

350

the use of Bayesian Belief Network (BBN) in mod-

elling agents’ knowledge. The selection of the learn-

ing algorithm depends on the nature of the data at

hand. For instance, counting algorithm ﬁts diagnos-

tic problems. But when the data contain latent vari-

ables, conjugate gradient descent or expectation max-

imisation algorithms could be the best options (Bot-

tou, 2010; Dempster et al., 1977; Romanycia, 2019)

3.3.4 Learning (C4)

The learning process is responsible for handling the

data, manage its uncertainty, and control the learn-

ing process. The uncertainty could be inputted to the

learning algorithms in one of the following ways (Ro-

manycia, 2019):

• Restricted or unrestricted range: in this approach,

agents can send ranges of knowledge of BBN val-

ues. For example, temperature = [250-300], i.e.,

the value of the temperature is between 250-300

or temperature >30 degrees Celsius.

• Possibility or impossibility list: Setting a list of

the possible values or negating the list to show im-

possibilities in these values. For example temper-

ature ={200, 250, 300}or temperature = ¬ {200,

250, 300}

• Likelihood: the set of likelihood probabilities can

be attached to the possible variables in restricted

or unrestricted form. For instance, temperature =

{200 .8, 250 .1+-1, 300 .1}.

• Complete or incomplete certainty: It happens

when the BBN has a complete doubt about the

variable, or it has no doubt on the variable by pro-

viding its value to the BBN as “?” or actual values.

The output of the learning process is a well-trained

BBN (i.e., if BBN were used in modelling the agent’s

knowledge) serve as the output C5.

3.3.5 Output (C5)

The output of the learning process could be an accu-

rate prediction tool (e.g., well-trained Bayesian Belief

Network, neuro-fuzzy system, etc.). The accuracy of

the prediction tool could be measured by considering

how many times the network predicted a wrong val-

ues known as prediction accuracy rate (Romanycia,

2019). It could be ranged between 0 and 1 with 0 be-

ing the best. The choice of the threshold depends on

the programmer’s choice and learning environment.

In case of highly changing environment, the learning

algorithms could be augmented to prioritised recent

data as in (Bottou, 2010; Romanycia, 2019).

The output network could be replicated to both

agent and human side as their predictive tool in syn-

chronous or asynchronous mode (as discussed in sec-

tion 3.3.1). The higher the mission data, the higher

the perfection of the learned BBN as tested in our ex-

periment. Therefore, this architecture is limited to the

availability of data. In order to solve that issue, we

propose the use of fuzzy logic (set of heuristic to mon-

itor the prediction (Dernoncourt, 2013)) or expert in-

put (to ﬁll in the conditional probability of the BBN)

in the absence of data.

4 IMPLEMENTATION OF THE

MODEL ON REAL-WORD

PROBLEM OF FOREST FIRE

LOOKOUT

We tested the model on the problem of wildﬁre

searching. Forest ﬁre is one of the world’s major

problems, it kills lots of human and animal lives, de-

stroy millions of acres of land, and affect the cli-

matic conditions (Ingle, 2011). We use a team of

multi-rotors and ﬁxed-wing UAVs mounted with ﬁre

detecting sensors (camera) simulated on Aerospace

Multi-agent Simulation Environment – AMASE (afrl

rq, 2019). The belief of the agent was modelled using

Bayesian Belief Network (BBN) on Netica (Romany-

cia, 2019). Each agent is updating its BBN, and the

data is recorded at the metadata section for training

purposes. We assume the structure in ﬁgure 2.

Figure 3 describes two quadrotors and two ﬁxed-

wing UAVs conducting forest ﬁre searching. The in-

set picture shows a human expert from the control

station room with a PC server and could communi-

cate with the agents. The human expert is capable of

seeing all the UAVs data, as described in ﬁgure 4 on

AMASE.

From ﬁgure 4, the places mark with alphabets de-

notes:

• A represents the UAVs.

• Places marked B and C are the ﬁres.

• D is the UAV’s information (speed, position,

heading, altitude, etc.) visible to the human ex-

pert via the PC connector at the base station..

• E is the sensor data of the UAVs.

• F is the battery level of the UAVs, and

• G is the environmental information sensor data

such as wind speed, wind direction, etc.

The agents could start with an in-built autonomous

searching approach. We use levy ﬂight of (Chawla

Human-agents Interactions in Multi-Agent Systems: A Case Study of Human-UAVs Team for Forest Fire Lookouts

351

Figure 3: Demonstration of Human-UAVs Team for Forest

Fire Searching.

Figure 4: Implementation of the Propose Architecture on

AMASE.

and Duhan, 2018) to monitor the UAVs waypoints

generations for searching activity. The agents’ belief

was modelled using Bayesian Belief Network on Net-

ica (Romanycia, 2019) as describe by ﬁgure 5.

Figure 5: Bayesian Belief Network for Monitoring the

UAVs Belief on Fire Presence.

Figure 5 describes the BBN for monitoring the UAVs

belief on ﬁre presence. Once the UAV’s sensor detects

the ﬁre, it will update the network and its metadata for

training purposes.

5 EXPERIMENTS AND RESULTS

This section describes the methodology and results

of each phase (stage of the model), implemented us-

ing AMASE (afrl rq, 2019) and Netica (Romanycia,

2019) on the forest ﬁre problem discuss in section 4.

All experiments were run on PC with 8GB RAM, in-

tel core i3-6006U @ 2GHZ, and I terabyte external

storage.

The clear idea of the process is the ability to obtain

a nice prediction tool that synchronises both agent and

human expert in the absence of communication. We

use conjugate gradient descent algorithm to train the

Spatio-temporal BBN (ﬁgure 6) to produce a possible

location and action of UAV at a given time. Therefore,

during communication, hardware, or software failure

the agents can continue with their mission, and the

human expert could be able to predict the possible lo-

cation and action of the agent at a given time.

Figure 6: Bayesian Belief Network for Monitoring the mis-

sion clock, agent’s location, and actions.

Figure 6 describes a simple BBN for monitoring the

agents’ actions, location, and time. The location

refers to segmented grids of small sizes say (2meters

square). Time class refers to the time range of the mis-

sion clock, e.g., 12:00-12:05. The actions of the agent

(e.g., searching, loitering, projection etc) at a particu-

lar time are recorded for training purposes. The BBN

simply says that, at every time t, the UAV has location

and action doing, this could be predicted by ﬁlling of

the BBN conditional probability values using the mis-

sion’s sensor data. If the BBN is well-trained (perhaps

with the prediction error rate of at most 0.1-0.3 out of

the worse 1, depending on the availability of training

data and nature of the environment), it could be used

to trace what action UAV is doing, in which location

at a particular time more especially in a static mission

by given the conditional probabilities obtained from

HAMT 2020 - Special Session on Human-centric Applications of Multi-agent Technologies

352

the learning process. The model could work in dy-

namic missions by prioritising the learning data as in

(Romanycia, 2019). The learning algorithms are also

capable of handling latent variables.

5.1 Testing the Model on Forest Fire

Searching

We tested the model using forest ﬁre searching prob-

lem introduced in section 4 using a team of UAVs

as agents simulated on Aerospace Multi-agent Sim-

ulation Environment -AMASE (afrl rq, 2019). In or-

der to compare the perfection of these techniques, we

tested both the agent’s BBN prediction error rates and

the human prediction perfection. The UAV is con-

tinually generating random waypoint using the bio-

inspired levy ﬂight searching technique of (Chawla

and Duhan, 2018). We evaluate the human expert part

by exposing ten volunteer participants to the system

and then monitor their guess accuracy on waypoint

locations at a particular time. Table 1 describes the

prediction error of the respective BBN.

Table 1: Prediction Error Rate Comparison between UAV

and Human Entry of BBN conditional Probability Table

Values.

BBN

Entry

Source

BBN

Pre-

diction

Error

Number of Training

Data

Human

Expert

0.815 Number of partici-

pants: 10

UAV 0.505 1000

UAV 0.303 2000

UAV 0.216 3000

UAV 0.166 4000

6 DISCUSSION

From the results in table 1 section 5, one can notice

that the UAVs learning prediction error rate is less

than the human error rate. This is unsurprising be-

cause the agents are generating their waypoints base

on a stochastic bio-inspired random approach(levy

ﬂight), and human could not be able to predict what

will happen next. However, the human expert entry

to the system is essential in terms of emergency com-

mands execution for the safety of the UAVs, as dis-

cussed in section 3. Another claim is that the predic-

tion error of the UAVs reduces with an increase in the

number of training data (table 1). The utilisation of

the training data to achieve the maximal outcome was

mark as future work.

6.1 Effect of the Model on Forest Fire

Searching

In terms of communication breakdown or (hardware

breakdown), the learned prediction error could be

used in making accurate predictions estimations, and

decision to control the mission. For example, let us

assume a power breakdown at the base station dur-

ing the human-UAV ﬁre searching describe in sec-

tion 4, the based station could be able to trace their

UAV energy level, failure location (in case it ﬁnishes

its battery), location, etc. This removes the neces-

sity of using communication in human-agent inter-

action popularly known as mixed-initiative reason-

ing and planning as in (Bevacqua et al., 2015; Ca-

cace et al., 2014; Ferguson and Allen, 2007; Makonin

et al., 2016; Tecuci et al., 2007). In terms of adapta-

tion, the BBN in ﬁgure 6 describes the way of adapt-

ing to the environment by the agent through the pro-

vision of conditional probabilities learning.

Therefore, ﬁnally, we here argue that accurate pre-

diction tools obtain via join human-machine learning

can help in monitoring mission during communica-

tion failure and enhance the adaptation in a human-

machine team.

7 CONCLUSIONS AND FUTURE

WORK

We proposed an architecture for monitoring the

human-agent team by utilising the best part of the

entities knowledge and producing accurate prediction

tools through the use of machine learning algorithms

(gradient descent or expectation maximisation). We

modelled the agents’ belief using Bayesian Belief

Network (BBN) and expose it to training data. In

order to test the proposed model, we used a forest

ﬁre monitoring by a team of UAVs and human ex-

pert in the base station. The human prediction proofs

to be inaccurate as expected but very useful in terms

of emergency control. We were able to get the pre-

diction accuracy of 0.166 by training the BBN us-

ing 4000 samples of the agent’s data. This is pretty

good for making estimation, predictions, and decision

in the absence of accurate communication. In terms

of the highly changing environment, we propose the

augmentation of the learning algorithms to prioritise

recent cases as in the fading strategy of (Romanycia,

2019).

In the future, we will introduce a clear strategy

Human-agents Interactions in Multi-Agent Systems: A Case Study of Human-UAVs Team for Forest Fire Lookouts

353

for the distributed learning process between the agent

and the human experts. This will propose a complete

concept of parallel learning. We are also planning to

optimise the learning data and dig deeper to explore

the nature of the prediction accuracy, and it is rele-

vant to the available data. Although our model intro-

duces faults tolerance and communication failure or

reduction, a comparative analysis with other systems

in (Makonin et al., 2016; Tecuci et al., 2007) using

real agents marked as future work. We will also look

at how the model and the predictions tools act in a

highly changing environment.

ACKNOWLEDGEMENTS

We appreciate the effort of Petroleum Technology

Trust Funds (PTDF) of Nigeria for sponsoring this

project.

REFERENCES

afrl rq (2019). https://github.com/afrl-rq/OpenAMASE.

afrl-rq. original-date: 2017-05-07T13:17:17Z.

Araragi, T. (2005). Fault tolerance for internet agent sys-

tems: In cases of stop failure and byzantine failure.

AAMAS ’05, page 123–124, New York, NY, USA.

ACM. event-place: The Netherlands.

Bevacqua, G., Cacace, J., Finzi, A., and Lippiello, V.

(2015). Mixed-initiative planning and execution

for multiple drones in search and rescue missions.

ICAPS’15, page 315–323. AAAI Press. event-place:

Jerusalem, Israel.

Bottou, L. (2010). Large-scale machine learning with

stochastic gradient descent. page 10.

Cacace, J., Finzi, A., and Lippiello, V. (2014). A mixed-

initiative control system for an aerial service vehi-

cle supported by force feedback. pages 1230–1235.

2014 IEEE/RSJ International Conference on Intelli-

gent Robots and Systems.

Chawla, M. and Duhan, M. (2018). Levy ﬂights in meta-

heuristics optimization algorithms – a review. Applied

Artiﬁcial Intelligence, 32(9-10):802–821.

Dempster, A. P., Laird, N. M., and Rubin, D. B. (1977).

Maximum likelihood from incomplete data via the em

algorithm. Journal of the Royal Statistical Society. Se-

ries B (Methodological), 39(1):1–38.

Dernoncourt, F. (2013). Introduction to fuzzy logic.

page 24.

Endert, A., Hossain, M. S., Ramakrishnan, N., North, C.,

Fiaux, P., and Andrews, C. (2014). The human is the

loop: New directions for visual analytics. J. Intell. Inf.

Syst., 43(3):411–435.

Endsley, M. R. (1995). Toward a theory of situation aware-

ness in dynamic systems. [Online; accessed 2019-11-

14].

Ferguson, G. and Allen, J. (2007). Mixed-initiative sys-

tems for collaborative problem solving. AI Magazine,

28(2):23–23.

Green, T. M., Maciejewski, R., and DiPaola, S. (2010). Al-

ida: Using machine learning for intent discernment in

visual analytics interfaces. pages 223–224. 2010 IEEE

Symposium on Visual Analytics Science and Technol-

ogy. ISSN: null.

Ingle, L. B. (2011). Every day is ﬁre day: A study of his-

toric ﬁre towers and lookout life in the great smoky

mountains national park. page 136.

Kifor, T., Varga, L. Z., Vazquez-Salceda, J., Alvarez, S.,

Willmott, S., Miles, S., and Moreau, L. (2006). Prove-

nance in agent-mediated healthcare systems. IEEE In-

telligent Systems, 21(6):38–46.

Makonin, S., McVeigh, D., Stuerzlinger, W., Tran, K., and

Popowich, F. (2016). Mixed-initiative for big data:

The intersection of human + visual analytics + pre-

diction. pages 1427–1436. 2016 49th Hawaii Interna-

tional Conference on System Sciences (HICSS).

Rich, C. and Sidner, C. L. (2007). Diamondhelp: A generic

collaborative task guidance system. AI Magazine,

28(2):33–33.

Rich, C., Sidner, C. L., and Lesh, N. (2001). Collagen:

Applying collaborative discourse theory to human-

computer interaction. AI Magazine, 22(4):15–15.

Romanycia, M. (2019). Netica-j reference manual. page

119.

Stanton, N. A., Stewart, R., Harris, D., Houghton, R. J.,

Baber, C., McMaster, R., Salmon, P., Hoyle, G.,

Walker, G., Young, M. S., Linsell, M., Dymott, R.,

and Green, D. (2006). Distributed situation awareness

in dynamic systems: theoretical development and ap-

plication of an ergonomics methodology. Ergonomics,

49(12-13):1288–1311.

Tecuci, G., Boicu, M., and Cox, M. T. (2007). Seven as-

pects of mixed-initiative reasoning:an introduction to

this special issue on mixed-initiative assistants. AI

Magazine, 28(2):11–11.

Turpin, M., Michael, N., and Kumar, V. (2014). Capt: Con-

current assignment and planning of trajectories for

multiple robots. The International Journal of Robotics

Research, 33(1):98–112.

Yue, J., Raja, A., and Ribarsky, W. (2010). Predictive ana-

lytics using a blackboard-based reasoning agent. vol-

ume 2, pages 97–100. 2010 IEEE/WIC/ACM Interna-

tional Conference on Web Intelligence and Intelligent

Agent Technology. ISSN: null.

HAMT 2020 - Special Session on Human-centric Applications of Multi-agent Technologies

354