Assessing Routing Decisions of Search and Rescue Teams in Service of an

Artiﬁcial Social Intelligence Agent

Volkan Ustun

1 a

, Ronit Jorvekar

, Nikolos Gurney

1 b

, David Pynadath

1,2 c

and Yunzhe Wang

3 d

Institute for Creative Technologies, University of Southern California, Playa Vista, CA, U.S.A.

Department of Computer Science, University of Southern California, Los Angeles, CA, U.S.A.

Department of Computer Science, Columbia University, New York, NY, U.S.A.

Keywords:

Artiﬁcial Social Intelligence, Graph Neural Networks, Route Planning, Urban Search and Rescue.

Abstract:

In the context of Urban Search and Rescue (USAR) missions, efﬁcient routing performance is of paramount

importance for the success of a USAR team. Artiﬁcial Social Intelligence (ASI) agents could play a cru-

cial role in guiding and interacting with these teams, and an analysis of the routing choices made by USAR

teams can offer valuable insights into their overall performance and provide guidance for interventions by

ASI agents. This study capitalizes on recent advancements in Graph Neural Networks, transformers, and at-

tention models to harness their capabilities as neural heuristics for rapidly generating near-optimal routes in

routing challenges. Speciﬁcally, we propose a real-time decision framework to scrutinize and evaluate rout-

ing decisions executed by participants during the DARPA ASIST Minecraft USAR Task. This assessment

involves comparing the routing decisions made by participants and routes concurrently generated and recom-

mended by neural heuristics employing Graph Neural Networks with attention mechanisms. Furthermore, our

investigation delves into the potential of routing decision assessments as informative indicators for an ASI

agent, aiding in identifying scenarios necessitating intervention. This research contributes to using quantita-

tive metrics, such as routing efﬁciency, as meaningful signals for ASI agents to monitor the performance of

USAR teams through integrating state-of-the-art AI techniques. Ultimately, this integration could enhance the

efﬁciency and effectiveness of an ASI in guiding search and rescue operations.

1 INTRODUCTION

Autonomous agents have the potential to enhance

teamwork by automatically assessing and providing

assistance during task performance (Sukthankar et al.,

2007; Webber et al., 2019). Furthering research in this

domain is the objective of DARPA’s ASIST program,

which employs a simulated urban search and rescue

(USAR) task using a Minecraft environment to eval-

uate autonomous agents (Freeman et al., 2023). This

task environment is utilized to run experiments with

distributed teams of three participants aided by an Ar-

tiﬁcial Social Intelligence (ASI) agent acting as an ad-

visor. The success and safety of the USAR team are

heavily reliant on routing decisions, necessitating the

tracking of participant and team routing decisions by

https://orcid.org/0000-0002-7090-4086

https://orcid.org/0000-0003-3479-2037

https://orcid.org/0000-0003-2452-4733

https://orcid.org/0000-0003-3862-1163

a capable USAR ASI agent.

A USAR ASI agent can use routing in multiple

ways. For instance, it can aid with real-time routing

decisions of search and rescue teams by utilizing the

outputs of a route generator resembling a navigation

app that creates routes and recommends alternatives

based on present data such as road conditions, trafﬁc

congestion, accidents, and more. However, humans

do not always strictly adhere to the recommendations

of a navigation app because they consider their own

private information, including preferences and their

physiological and emotional states. To illustrate, a

driver with a car full of hungry kids may choose to

stop and eat dinner early after the app recommends

a re-route rather than following the recommendation.

Similar to how the app performs real-time path plan-

ning and offers choices to the user based on current

data, a USAR ASI agent can ingest routing sugges-

tions and merge them with other mission-related in-

formation and the affective state of the USAR team to

provide more effective guidance.

Ustun, V., Jorvekar, R., Gurney, N., Pynadath, D. and Wang, Y.

Assessing Routing Decisions of Search and Rescue Teams in Service of an Artiﬁcial Social Intelligence Agent.

DOI: 10.5220/0012388100003636

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 16th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2024) - Volume 1, pages 313-320

ISBN: 978-989-758-680-4; ISSN: 2184-433X

313

Another example of how an ASI agent can lever-

age routing is to monitor the team’s routing decisions

to assess the team’s overall state. For instance, a de-

cline in the quality of routing decisions may indicate

potential issues within the search and rescue team or

suggest that the team is not fully leveraging avail-

able information. The ASI agent can use these ob-

servations and inferences to provide improved sug-

gestions and interventions, ultimately enhancing the

team’s performance.

Generating good routing options in real-time is

critical for both use cases mentioned above, as they

require considering the current state of the USAR

mission. However, this task is challenging because

ASIST’s USAR routing is similar to the family of NP-

Hard Traveling Salesman Problems (TSPs). As a re-

sult, heuristics must be employed to ﬁnd candidate

solutions, as is the case with many practical appli-

cations (Boussa

ıd et al., 2013). One helpful insight

for developing heuristics for routing problems is that

the problem instances often share common character-

istics or patterns, as demonstrated by (Cappart et al.,

2021) in a trucking company routing problem where

the company needs to generate daily routes for the

same city with slight variations due to trafﬁc con-

ditions. These similarities provide opportunities for

data-dependent machine learning approaches that can

exploit common patterns (Bengio et al., 2021). In re-

cent years, Graph Neural Networks (GNNs) with at-

tention mechanisms have emerged as effective heuris-

tic alternatives for combinatorial optimization prob-

lems (Cappart et al., 2021). Leveraging such an ap-

proach as a neural heuristic can rapidly generate good

paths that utilize the similarities in routing require-

ments of a Minecraft USAR task. A real-time ASI

agent prototype can take advantage of this capability

to explore the routing options available to the SAR

team under varying conditions.

(Wang et al., 2023) discusses our preliminary ex-

plorations on how GNNs with attention mechanisms

can be leveraged as neural heuristics to generate good

routes for particular states of Minecraft USAR tasks.

We build on this earlier work and present a more capa-

ble pipeline that utilizes more expressive embeddings

and higher-quality training data, resulting in 11% im-

provement in the length of the generated routes. Fur-

thermore, we perform a new exploratory analysis of

the data collected on participant teams performing the

Minecraft USAR task and discuss how scrutinizing

routing performance can guide the timing of the in-

terventions provided by an ASI agent.

2 BACKGROUND

Combinatorial optimization (CO) is an established

interdisciplinary ﬁeld that has numerous real-world

applications, including routing (Korte and Vygen,

2012). Its primary goal is to optimize a cost or

objective function by selecting a subset from a ﬁ-

nite set while adhering to selection constraints. CO

strives to obtain a unique and optimal solution for

each problem, but the complexity of certain problems

can make this impractical. In such cases, practition-

ers often rely on problem-speciﬁc heuristic method-

ologies (Boussa

ıd et al., 2013). However, practi-

cal situations frequently involve problem instances

that share speciﬁc characteristics or patterns (Cappart

et al., 2021). These similarities present opportunities

for data-dependent machine-learning approaches that

can leverage these patterns (Bengio et al., 2021). For

example, (Cappart et al., 2021)cite the example of

a trucking company’s daily routing solutions for the

same city, with slight variations in travel times due to

varying trafﬁc conditions.

Graph Neural Networks (GNNs) are a potent ma-

chine learning architecture that exploits structural, re-

lational, and compositional biases to facilitate geo-

metric deep learning (Gilmer et al., 2017). GNNs ag-

gregate information from structural and feature-based

(e.g., node or edge type) graph data into simpler rep-

resentations of nodes and edges. By parameteriz-

ing this aggregation, they can be trained end-to-end

against a loss function. GNNs can operate on higher

complexity data than what can be represented in regu-

lar Euclidean structures, such as an image (2D) or text

(1D). GNNs achieve this by being order-invariant,

propagating on each node in the graph independently,

ignoring the input order, and using the graph structure

to guide propagation. These innovations empower

GNN models to ”reason” about a graph, make gen-

eral inferences, and use those inferences to make pre-

dictions and classiﬁcations successfully (Zhou et al.,

2020). In recent years, GNNs have been used as neu-

ral heuristics to generate solutions for CO problems

(Vesselinova et al., 2020). The primary promise of

GNNs in this role is that the learned vector represen-

tations encode critical graph structures to help solve

CO problems more efﬁciently (Cappart et al., 2021).

In 2018, (Kool et al., 2018) proposed a

transformer-like encoder-decoder architecture based

on Graph Attention Networks (Veli

ckovi

c et al.,

2017) for general routing problems. Their approach

trained an encoder-decoder neural network using an

actor-critic reinforcement learning approach on ran-

domly generated routing problems. The training did

not require optimal solutions to the training instances

ICAART 2024 - 16th International Conference on Agents and Artiﬁcial Intelligence

314

and could be done in advance. With the trained

model, it was possible to generate high-quality so-

lutions to SAR routing problems quickly. Recently,

we utilized Kool et al.’s codebase (2018) and aug-

mented it to generate routing solutions for an Arti-

ﬁcial Social Intelligence (ASI) agent (Wang et al.,

2023). Our task involved a capacitated vehicle rout-

ing with-proﬁts model that mapped to the tasks and

roles of human participants in a Minecraft USAR mis-

sion. Using the trained models, we were able to gen-

erate reasonable solutions quickly, informing the ASI

agent of potentially good solutions for the given mis-

sion state.

2.1 Minecraft USAR Test Environment

One of the main obstacles in creating an agent that can

aid human teams is assessing its performance. Prior

studies, such as the Electric Elves (Chalupsky et al.,

2002), demonstrated the potential of deploying such

agents in a real-world research lab. However, they

did not quantitatively evaluate their impact on team-

work in the traditional social science sense. There-

fore, a signiﬁcant accomplishment of the ASIST pro-

gram has been the development of a robust testbed

implemented in a game-based (Minecraft) environ-

ment, which enables distributed teaming tasks and al-

lows for thorough quantitative evaluations (Freeman

et al., 2023). Our approach is to model players, the

team, and team processes within the Minecraft en-

vironment as Partially Observable Markov Decision

Processes (POMDPs). Moreover, our agent incorpo-

rates input from analytical components of the testbed

as sensors to measure the team and the impact of in-

terventions. Our agent uses behavioral data to inform

these POMDP models through Inverse Reinforcement

Learning (IRL) and then uses these POMDPs recur-

sively to form a Theory of Mind about the team to

reason about the expected utility of alternate interven-

tions. (Pynadath et al., 2023)

The DARPA ASIST Minecraft SAR task environ-

ment (Figure 1) is an immersive training platform de-

signed to create Artiﬁcial Social Intelligence (ASI)

agents. The game’s objective is to rescue victims

of an urban disaster while earning points. The vic-

tims can be either non-critical or critical, and criti-

cal victims require a coordinated effort to save. The

game environment includes various challenges, such

as risks, where a player can get trapped and need as-

sistance from teammates, and rubble, which can im-

pede access to victims. Teams consist of three play-

ers who can choose from three roles: medical special-

ist (medic), hazardous material specialist (engineer),

and search specialist (transporter). The medics can

triage victims and rescue frozen teammates, the engi-

neers can clear rubble, and the transporters can trans-

port victims. The ASI agents are not embodied team

members. Still, they can communicate via text chat

with human team members (who can communicate

through a shared audio channel).

3 APPROACH

In the Minecraft USAR environment, participants are

required to perform a range of tasks across different

locations. The primary objective of this environment

is to triage victims and move them to secure areas,

with the order of victim triage being crucial for mis-

sion success due to time constraints. Therefore, the

medic role needs to focus on minimizing travel dis-

tance, which is similar to the Traveling Salesman’s

Problem, an NP-Hard optimization problem. How-

ever, in this setting, not all relevant locations are

known in advance, making it impossible to create an

optimal tour before the mission commences. Instead,

the ASI agent requires a general framework to quickly

generate satisfactory solutions to support or monitor

participant performance.

As mentioned, previous work in the ASIST pro-

gram (Wang et al., 2023) builds on the codebase

developed by (Kool et al., 2018) and presents a

framework that deﬁnes a semantic graph based on

Minecraft USAR task maps. This semantic graph

captures all the main map entities and structures, in-

cluding rooms, connections between rooms, victims,

and rubble locations, among other features that could

be utilized in navigation decisions as depicted in Fig-

ure 2. In this representation, each role has different

objective nodes; for example, the medic role utilizes

the victim nodes, whereas the engineer role requires

the rubble nodes that block the victims. To calculate

distances between nodes of interest, such as victims,

we use Dijkstra’s algorithm (Dijkstra, 1959), which

considers the layout of the environment. The orig-

inal framework includes a pipeline to extract infor-

mation from the maps and convert the resulting dis-

tance matrix into a set of 2D coordinates ﬁrst using

Metric Multidimensional Scaling (mMDS) (Kruskal,

1978; Cox and Cox, 2008), also known as Princi-

pal Coordinate Analysis (PCoA), and then the John-

Lindenstrauss Transform (JLT) (Johnson and Linden-

strauss, 1984) while preserving distance information

from the original distance matrix (Indyk et al., 2017).

The 2D coordinates are then scaled to [0, 1] to ﬁt the

original codebase’s requirements. The previous work

used neural heuristics for route planning on this 2D

coordinate set and converted the resulting routes back

Assessing Routing Decisions of Search and Rescue Teams in Service of an Artiﬁcial Social Intelligence Agent

315

Figure 1: The participant’s (medic) interface while playing the Minecraft USAR Scenario.

to the semantic map representation for visualization,

analysis, and interfacing purposes.

Our updated framework removes the ﬁnal step in

the pipeline and uses the coordinates in 36D space

output by mMDS directly in route planning neural

heuristics. This alteration to the pipeline allows for

better preservation of the original distance matrix, as

converting back to 2D space was causing a loss of pre-

cision. The 36D coordinates serve as initial node em-

beddings for GNN representations. Additionally, we

experimented with including victim type as an addi-

tional node embedding, given the importance of dis-

tinguishing between critical and regular victims for

triage.

We simpliﬁed the roles of the transporter and en-

gineer by assuming they would follow the medic and

perform tasks as required. For example, the trans-

porter transports victims and assists in triaging critical

victims while the engineer clears rubble that blocks

victims or exits in threat rooms. While a more com-

plex logic could yield potential gains, a simple heuris-

tic in which the engineer and transporter follow the

medic was sufﬁcient to generate routes to complete

the mission on time. This simpliﬁcation also stream-

lines GNN model training by using the total distance

traveled by the medic as the primary cost/objective

function.

To train the GNN models, we generated 7,000 se-

mantic maps, similar to the number of training in-

stances used in our preliminary explorations, con-

structed similarly to the Minecraft USAR layouts,

with variations in victim locations, threat rooms, and

rubble blockages. In addition, we deﬁned ten addi-

tional semantic maps as test instances, two of which

are the original maps used in the Minecraft USAR

task environment.

4 EXPERIMENTS

We trained our model(s) with an actor-critic reinforce-

ment learning approach on the 7,000 randomly gen-

erated semantic maps on the base layout. With this

approach, we did not supervise the training with the

optimal solutions but rather generated gradients based

on the best-known solution found so far. We can uti-

lize the trained models in two different ways: (1) As a

route generator, where the ASI agent could utilize the

model to suggest routes to the team members, and (2)

As a team tracker, where the ASI agent could utilize

the model to track the team performance with respect

to the near-optimal routes generated by the trained

model.

4.1 Route Generation

The success of Minecraft USAR missions heavily re-

lies on effective navigation in the Minecraft envi-

ronment. When the locations of victims are known

in advance, minimizing the total mission time for

ICAART 2024 - 16th International Conference on Agents and Artiﬁcial Intelligence

316

Figure 2: Minecraft USAR Task Map for the Saturn layout and the corresponding Semantic Graph.

the medic is comparable to solving a TSP problem.

Therefore, if an ASI agent can access near-optimal

routes, it can use them to suggest victim triage plans.

We applied this approach to compare the quality of

solutions for different conﬁgurations, as shown in

Table 1. Although all three conﬁgurations produce

routes that complete the mission within the 1020-

second time limit, directly utilizing the 36D node em-

beddings signiﬁcantly improves the solution quality,

as expected. Moreover, adding victim type as an addi-

tional node embedding to our representation leads to

slight improvements in solution quality since utilizing

victim type information could enhance synchroniza-

tion among the participants. Finally, the two ground

truth layouts used in the Minecraft USAR experi-

ments have solution times of 765 and 770 seconds,

respectively, which are well below the 1020-second

mark.

Table 1: Mission completion times in seconds for the test

set instances.

Model Average Median St. Dev

2D 770.06 767.34 42.16

36D 699 699.12 41.32

36D + Type 685.62 682.66 34.24

Although victim locations may be known in ad-

vance, planning the route from scratch may not al-

ways be feasible. The team may do things differently

than the suggested routes, and hence, the route gen-

eration process must consider various starting states.

Our solution pipeline can accommodate such changes

in initializations. For instance, in a Human Subject

Study, we demonstrated that we could run our solu-

tion pipeline from any point during the experiment,

given the state of the Minecraft USAR task. Fur-

thermore, the generated routing suggestions could en-

hance the team’s performance for the remainder of the

mission compared to what had happened in the ac-

tual mission. In a particular experiment, for example,

a human subject team’s performance in scored game

points could have improved by 30 % if they followed

the routing suggestion generated by our pipeline for

the rest of the mission at the 2/3rd mark.

However, in both the Minecraft USAR environ-

ment and real-life USAR missions, it is typically not

feasible to know the locations of victims in advance.

Therefore, the actual use case is more complex, even

with real-time near-optimal route planning capabil-

ity. For instance, in the Minecraft USAR environ-

ment, each participant is provided unique knowledge

based on their roles, including information about po-

tential victim locations, threat rooms, sensory cues to

assist in ﬁnding victims, and more. In Human Subject

Experiments, participants begin the scenario with this

knowledge and simultaneously execute the assigned

search and rescue task while exploring the environ-

ment. When we initiated our solution pipeline with

the same knowledge and only planned for the known

artifacts (but perceived additional information during

the execution of the planned routes), the model lo-

cated and triaged 85% of the victims in around 660

seconds, leaving enough time for further exploration.

However, we did not implement an exploration algo-

rithm to augment our model for this task version.

4.2 Team Tracking

An ASI agent participating in the DARPA ASIST pro-

gram is expected to facilitate better teamwork through

interventions. Although there are various approaches

to making such interventions, monitoring the naviga-

tion performance of a team and identifying signiﬁcant

shifts in this performance can be a valuable indicator

for the ASI agent. In the Minecraft USAR environ-

ment described by (Freeman et al., 2023), a Human

Subject Research study was carried out to evaluate the

performance of ASI agents. The study involved 242

trials with different conﬁgurations, and our analysis

focuses on 238 of these trials (due to data issues with

the remaining 4). Our investigation aims to determine

Assessing Routing Decisions of Search and Rescue Teams in Service of an Artiﬁcial Social Intelligence Agent

317

whether the neural heuristics’ near-optimal routes can

be utilized to track the navigation performance of hu-

man teams.

Each trial includes a metadata ﬁle that captures

the Minecraft environment state during the mission.

At mission times 0, 3, 5, 8, and 11 minutes, we took

snapshots of the environment state and used our so-

lution pipeline to generate solutions for the remain-

ing tasks based on that speciﬁc state. It is important

to note that our solution pipeline only focuses on un-

ﬁnished tasks and does not attempt to complete tasks

that have already been completed, such as triaging

a victim. This approach allowed us to gather a list

of lower bounds on mission completion times based

on the progress made by the team up to each speciﬁc

mission time. We then used these hypothetical com-

pletion times to calculate the deviation from the near-

optimal route that was calculated at time 0, which rep-

resents a complete solution generated with perfect in-

formation. By tracking these deviations over time, we

can detect signiﬁcant changes in the team’s navigation

performance and identify opportunities for interven-

tion.

Figure 3: Multivariable linear regression analysis of percent

deviations from the optimal route.

To analyze the data, we used multivariable linear

regression, and even with the inherent conﬁguration

variations in each trial, there is a trendline as shown

in Figure 3. We expect that for most teams, the devi-

ation from the optimal route will increase over time,

following the trendline. However, major deviations

from this trendline would indicate a signiﬁcant shift

in the team’s navigation performance, which would

warrant intervention.

To identify major deviations from the expected

trendline, we can examine the standard residual plots

of the multivariable linear regression model, as shown

in Figure 4. In particular, any standard residual above

2 (represented by the red dots on the plot) would war-

rant further investigation to determine whether an in-

tervention is needed to improve the team’s perfor-

mance. Conversely, standard residuals less than -

2 (represented by the green dots on the plot) could

indicate areas where the team’s navigation perfor-

mance can be improved, providing valuable insights

into their overall performance.

Figure 4: Standard residuals from the multivariable linear

regression model of deviations from optimal route.

5 DISCUSSION

We have signiﬁcantly improved the existing solu-

tion pipeline, resulting in better solution quality and

demonstrating its potential as a reliable tool for a

USAR ASI agent. However, it’s important to note that

the pipeline’s performance superior to human par-

ticipant teams was expected to some extent due to

its access to maps and victim locations. Nonethe-

less, Graph Neural Networks offer a robust architec-

ture for developing neural heuristics for combinatorial

problems like routing, particularly when there is sig-

niﬁcant similarity among problem instances. Graph

Neural Networks’ ﬂexibility and ability to adapt to

changes in objective functions make them an appro-

priate choice for handling additional problem infor-

mation through node embeddings. This ﬂexibility,

along with the ability to rapidly generate routing so-

lutions under different conditions, including starting

with limited information similar to human participant

teams, allows the ASI agent to better evaluate the

routes available to a USAR team.

To speed up training while still generating qual-

ity solutions, we made the assumption in our solution

pipeline that the engineer and the transporter would

follow the medic, as we found that using a more in-

sightful heuristic for routing them did not result in any

improvements in our exploratory experiments. How-

ever, we acknowledge that this assumption may not

be the most effective for an actual deployment, and

further improvements may be necessary to optimize

the use of the transporter and the engineer.

Although we believed that monitoring the naviga-

ICAART 2024 - 16th International Conference on Agents and Artiﬁcial Intelligence

318

tion performance of participants and detecting devi-

ations would provide valuable insights for the ASI

agent on the state of the teams, we were unable to

test these ideas due to budget cuts. Furthermore, in

our exploratory analyses, we did not have enough in-

formation to determine whether teams with high stan-

dard residual values had actual problems. Thus, our

results are mainly exploratory, and further investiga-

tion is needed to assess the effectiveness of our pro-

posed approach.

6 CONCLUSIONS

We have developed a versatile routing system that uti-

lizes neural heuristics to efﬁciently guide a real-time

ASI agent on available routing options for a USAR

team based on the current state of the mission. This

system can serve as a reliable tool for the ASI agent

to analyze routing options for the USAR team it is as-

sisting. Additionally, the framework enables the ASI

agent to monitor the team’s navigation performance

and identify any potential difﬁculties they may be ex-

periencing. By leveraging this information along with

other insights, detecting such issues can prompt effec-

tive interventions by the ASI agent.

ACKNOWLEDGEMENTS

Part of the effort depicted is sponsored by the U.S.

Army Research Laboratory (ARL) under contract

number W911NF-14-D-0005 and by the Defense Ad-

vanced Research Projects Agency (DARPA) under

contract number W911NF2010011, and that the con-

tent of the information does not necessarily reﬂect the

position or the policy of the Government or the De-

fense Advanced Research Projects Agency, and no of-

ﬁcial endorsements should be inferred.

REFERENCES

Bengio, Y., Lodi, A., and Prouvost, A. (2021). Machine

learning for combinatorial optimization: a method-

ological tour d’horizon. European Journal of Oper-

ational Research, 290(2):405–421.

Boussa

ıd, I., Lepagnot, J., and Siarry, P. (2013). A survey

on optimization metaheuristics. Information sciences,

237:82–117.

Cappart, Q., Ch

etelat, D., Khalil, E., Lodi, A., Morris, C.,

and Veli

ckovi

c, P. (2021). Combinatorial optimiza-

tion and reasoning with graph neural networks. arXiv

preprint arXiv:2102.09544.

Chalupsky, H., Gil, Y., Knoblock, C. A., Lerman, K., Oh, J.,

Pynadath, D. V., Russ, T. A., and Tambe, M. (2002).

Electric elves: Agent technology for supporting hu-

man organizations. AI magazine, 23(2):11–11.

Cox, M. A. and Cox, T. F. (2008). Multidimensional scal-

ing. In Handbook of data visualization, pages 315–

347. Springer.

Dijkstra, E. W. (1959). A note on two problems in connex-

ion with graphs. Numerische mathematik, 1(1):269–

271.

Freeman, J., Huang, L., Wood, M., and Cauffman, S. J.

(2023). Evaluating artiﬁcial social intelligence in an

urban search and rescue task environment. In Compu-

tational Theory of Mind for Human-Machine Teams:

First International Symposium, ToM for Teams 2021,

Virtual Event, November 4–6, 2021, Revised Selected

Papers, pages 72–84. Springer.

Gilmer, J., Schoenholz, S. S., Riley, P. F., Vinyals, O., and

Dahl, G. E. (2017). Neural message passing for quan-

tum chemistry. In International conference on ma-

chine learning, pages 1263–1272. PMLR.

Indyk, P., Matou

sek, J., and Sidiropoulos, A. (2017). 8:

low-distortion embeddings of ﬁnite metric spaces. In

Handbook of discrete and computational geometry,

pages 211–231. Chapman and Hall/CRC.

Johnson, W. B. and Lindenstrauss, J. (1984). Extensions of

lipschitz mappings into a hilbert space 26. Contempo-

rary mathematics, 26.

Kool, W., Van Hoof, H., and Welling, M. (2018). Atten-

tion, learn to solve routing problems! arXiv preprint

arXiv:1803.08475.

Korte, B. and Vygen, J. (2012). The traveling salesman

problem. In Combinatorial Optimization, pages 557–

592. Springer.

Kruskal, J. B. (1978). Multidimensional scaling. Num-

ber 11. Sage.

Pynadath, D., Gurney, N., Kenny, S., Kumar, R., Marsella,

S., Matuszak, H., Mostafa, H., Sequeira, P., Ustun, V.,

and Wu, P. (2023). Effectiveness of teamwork-level

interventions through decision-theoretic reasoning in

a minecraft search-and-rescue task. In Proceedings

of the 2023 international conference on autonomous

agents & multiagent systems.

Sukthankar, G., Sycara, K., Giampapa, J. A., Burnett, C.,

and Preece, A. (2007). Towards a model of agent-

assisted team search. In Procceedings of the First An-

nual Conference of the International Technology Al-

liance in Network and Information Science.

Veli

ckovi

c, P., Cucurull, G., Casanova, A., Romero, A., Lio,

P., and Bengio, Y. (2017). Graph attention networks.

arXiv preprint arXiv:1710.10903.

Vesselinova, N., Steinert, R., Perez-Ramirez, D. F., and Bo-

man, M. (2020). Learning combinatorial optimization

on graphs: A survey with applications to networking.

IEEE Access, 8:120388–120416.

Wang, Y., Gurney, N., Zhou, J., Pynadath, D. V., and

Ustun, V. (2023). Route optimization in service

of a search and rescue artiﬁcial social intelligence

agent. In Computational Theory of Mind for Human-

Machine Teams: First International Symposium, ToM

Assessing Routing Decisions of Search and Rescue Teams in Service of an Artiﬁcial Social Intelligence Agent

319

for Teams 2021, Virtual Event, November 4–6, 2021,

Revised Selected Papers, pages 220–228. Springer.

Webber, S. S., Detjen, J., MacLean, T. L., and Thomas, D.

(2019). Team challenges: Is artiﬁcial intelligence the

solution? Business Horizons, 62(6):741–750.

Zhou, J., Cui, G., Hu, S., Zhang, Z., Yang, C., Liu, Z.,

Wang, L., Li, C., and Sun, M. (2020). Graph neu-

ral networks: A review of methods and applications.

AI Open, 1:57–81.

ICAART 2024 - 16th International Conference on Agents and Artiﬁcial Intelligence

320