Prediction-Based Selective Negotiation for Refining
Multi-Agent Resource Allocation
Madalina Croitoru
1
, Cornelius Croitoru
2
and Gowrishankar Ganesh
3
1
University of Montpellier, LIRMM, France
2
Faculty of Computer Science, Al. I. Cuza Univerisity, Iasi, Romania
3
LIRMM, CNRS (National Center of Scientific Research), France
Keywords:
Collective Decision Making, Multi Agent Systems, Negotiation.
Abstract:
This paper proposes a 2-stage framework for multi-agent resource allocation. Following a Borda-based al-
location, machine learning predictions about agent preferences are used to selectively choose agent pairs to
perform negotiations to swap resources. We show that this selective negotiation improves overall satisfaction
towards the resource redistribution.
1 INTRODUCTION
Resource allocation (Ibaraki and Katoh, 1988) is a
core problem in a wide range of real-world multi-
agent applications (Chevaleyre et al., 2005), from
distributing medical supplies in emergency situations
(Zhang et al., 2016) to assigning computational re-
sources in cloud computing (Vinothina et al., 2012)
or allocating advertising slots to bidders in digital
marketplaces (Li et al., 2018). Such problems re-
quire dividing limited resources among competing
agents, each with their own preferences. Finding opti-
mal solutions to such problems is typically NP-hard,
with linear programming methods used for approxi-
mate solutions (Katoh and Ibaraki, 1998; Croitoru and
Croitoru, 2011).
Advancements in generative Artificial Intelligence
(AI) and autonomous systems are shaping hybrid so-
cieties where human agents and artificial systems are
deeply interconnected. In these environments, sub-
jective, context-dependent human preferences must
integrate with the pre-coded preferences of artificial
agents. Resource allocation problems involving hu-
man agents should consider contextual factors such as
socioeconomic status or geographic location, as these
can influence priorities that may evolve during the al-
location process (Dafoe et al., 2020). For instance,
in digital marketplaces, preferences for goods or ser-
vices can be inferred from browsing behavior or de-
mographic profiles. Similarly, in education, predic-
tive models can guide resource allocation to students
based on their specific learning needs.
In hybrid societies, achieving the global good re-
quires systems that facilitate fair trade-offs among di-
verse stakeholders, balancing individual preferences
with collective outcomes. This paper examines a
straightforward approach to negotiation: the swap-
ping of goods. This method serves as a form
of compromise, enabling mutually acceptable out-
comes through direct exchanges while supporting the
broader collective interest. We propose a 2-stage
framework refining multi-agent resource allocation:
1. In the first step, individual preferences are ag-
gregated using the Borda voting method to cre-
ate a collective preference ranking. This aggre-
gated ranking, combined with the individual pref-
erences of each agent, is used to allocate goods in
a greedy manner. Starting with the most preferred
good in the Borda aggregated ranking, goods are
allocated to the maximum number of agents, pro-
ceeding sequentially to less preferred goods.
2. In the second step, following a Condorcet-like ap-
proach, pairs of agents are identified based on ma-
chine learning predictions of their features, indi-
cating potential for compromise. These agents
are then considered for swapping their allocated
goods to enhance overall societal satisfaction by
aligning the allocations with predicted prefer-
ences.
In this work, our contributions are threefold:
• We formalize the 2-stage framework for
prediction-based resource allocation.
• We design an algorithm that integrates preference
prediction and preference aggregation for overall
agent satisfaction.
656
Croitoru, M., Croitoru, C. and Ganesh, G.
Prediction-Based Selective Negotiation for Refining Multi-Agent Resource Allocation.
DOI: 10.5220/0013368000003890
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 17th International Conference on Agents and Artificial Intelligence (ICAART 2025) - Volume 1, pages 656-662
ISBN: 978-989-758-737-5; ISSN: 2184-433X
Proceedings Copyright © 2025 by SCITEPRESS – Science and Technology Publications, Lda.

• Third, we analyse the satisfaction guarantees pro-
vided by the proposed framework.
The paper is structured as follows. Section 2 pro-
vides a motivating example illustrating how the use
of swaps can improve outcomes compared to a sim-
ple naive allocation. The naive allocation relies on
a lexicographical ordering of agents, assigning each
their most preferred goods in sequence until all goods
are allocated. Section 3 formally defines the resource
allocation problem and introduces the pseudocode for
the naive algorithm used in the motivating example
presented in Section 2. Section 4 details the Borda
voting method, which serves as the basis for the re-
fined allocation process described in the paper, and
the Condorcet voting method, which relies on pair-
wise comparisons of preferences. The intuition be-
hind Condorcet’s pairwise comparisons is used to jus-
tify why agents might swap goods. Section 5 intro-
duces our two-step framework and its theoretical sat-
isfaction guarantees. The framework (i) uses Borda
aggregation for an initial allocation, followed by (ii)
optimization through swaps between agents, guided
by feature-based preference predictions. Section 6
concludes the paper.
2 MOTIVATING EXAMPLE
This section shows an illustrative example involving
10 agents and 5 goods, each with a maximum avail-
ability of 4 units. The scenario explains how a simple
greedy initial allocation, while respecting multiplicity
constraints, may fail to achieve optimal satisfaction.
Using predicted preferences provided by a machine
learning model, we are then able to target the agents
that might engage in negotiation to improve the allo-
cation by augmenting overall satisfaction.
We consider 10 children A = {a
1
,a
2
,. . . ,a
10
} and
a pool of 5 goods G = {g
1
,g
2
,g
3
,g
4
,g
5
}, each with a
multiplicity of 4. Each child has a valuation function
v
i
: G → {15,10,5,0}, representing their preferences
for the goods. The top three preferences for each child
are shown in Table 1.
An initial allocation respecting multiplicities (no
good allocated to more than 4 children) is:
O
1
= {g
1
}, O
2
= {g
1
}, O
3
= {g
1
}, O
4
= {g
1
},
O
5
= {g
2
}, O
6
= {g
2
}, O
7
= {g
3
}, O
8
= {g
3
},
O
9
= {g
4
}, O
10
= {g
5
}.
The overall initial allocation satisfaction is:
v
A
(O) =
10
∑
i=1
v
i
(O
i
)
= 15 + 15 + 15 + 15 + 10 + 15 + 15 + 10
+ 15 + 0 = 135.
Using negotiations children can adjust the alloca-
tion to improve overall satisfaction:
• a
5
, receiving g
2
(value 10), negotiates with a
10
to
swap g
2
for g
1
, as g
1
is a
5
’s top preference and g
2
is a
10
’s second preference.
• a
10
, currently receiving g
5
(value 0), negotiates
with a
7
to swap g
5
for g
3
, as g
3
is a
10
’s top pref-
erence and g
5
is in a
7
’s top three.
After negotiation, the final allocation is:
O
∗
1
= {g
1
}, O
∗
2
= {g
1
}, O
∗
3
= {g
1
},
O
∗
4
= {g
1
}, O
∗
5
= {g
1
},O
∗
6
= {g
2
}, O
∗
7
= {g
2
},
O
∗
8
= {g
3
}, O
∗
9
= {g
4
}, O
∗
10
= {g
3
}.
The overall final allocation satisfaction is:
v
A
(O
∗
) =
10
∑
i=1
v
i
(O
∗
i
) = 150.
3 RESOURCE ALLOCATION
A resource allocation problem is defined as a tuple
(A, G, M,V ), where:
• A = {a
1
,a
2
,. . . ,a
n
} is the set of n ≥ 2 agents.
• G = {g
1
,g
2
,. . . ,g
m
} is the set of m ≥ 1 goods.
• M = {m
1
,m
2
,. . . ,m
m
} is the multiplicity m
j
≥ 1
of each good g
j
∈ G, representing the maximum
number of agents that can receive g
j
. The total
availability of goods is
∑
m
j=1
m
j
.
• V = {v
1
,v
2
,. . . ,v
n
} is the set of ordinal preference
functions, one for each agent a
i
∈ A. Each v
i
de-
fines a strict ranking over the goods G, such that
for any two goods g
j
,g
k
∈ G, v
i
(g
j
) < v
i
(g
k
) im-
plies that g
j
is strictly preferred to g
k
by agent a
i
.
An allocation O = {O
1
,O
2
,. . . ,O
n
} is a mapping
of goods to agents, where O
i
⊆ G denotes the subset
of goods allocated to agent a
i
. The allocation must
satisfy the multiplicity constraints:
∑
a
i
∈A
1
g
j
∈O
i
≤ m
j
, ∀g
j
∈ G,
where 1
g
j
∈O
i
is an indicator function that equals 1 if
g
j
∈ O
i
and 0 otherwise.
The (individual) satisfaction of agent a
i
is a func-
tion S
i
that associates to each subset G
′
⊆ G of goods
Prediction-Based Selective Negotiation for Refining Multi-Agent Resource Allocation
657

Table 1: Top 3 goods and their valuations for each child.
Child Top 3 Goods (Values v
i
(g))
a
1
v
1
(g
1
) = 15, v
1
(g
2
) = 10, v
1
(g
3
) = 5
a
2
v
2
(g
1
) = 15, v
2
(g
3
) = 10, v
2
(g
4
) = 5
a
3
v
3
(g
1
) = 15, v
3
(g
4
) = 10, v
3
(g
5
) = 5
a
4
v
4
(g
1
) = 15, v
4
(g
5
) = 10, v
4
(g
2
) = 5
a
5
v
5
(g
1
) = 15, v
5
(g
2
) = 10, v
5
(g
3
) = 5
a
6
v
6
(g
1
) = 15, v
6
(g
3
) = 10, v
6
(g
4
) = 5
a
7
v
7
(g
1
) = 15, v
7
(g
4
) = 10, v
7
(g
5
) = 5
a
8
v
8
(g
2
) = 15, v
8
(g
3
) = 10, v
8
(g
4
) = 5
a
9
v
9
(g
2
) = 15, v
9
(g
3
) = 10, v
9
(g
5
) = 5
a
10
v
10
(g
3
) = 15, v
10
(g
4
) = 10, v
10
(g
1
) = 5
a positive real value S
i
(G
′
). This value is strongly de-
pendent of the preferences values v
i
(g), for g ∈ G
′
.
The satisfaction of an allocation O is the sum of the
satisfactions of the agents for their respective alloca-
tions: S(O) =
∑
n
i=1
S
i
(O
i
). The objective is to find
an allocation O
∗
that maximizes satisfaction over all
possible allocations.
A straightforward allocation algorithm can allo-
cate goods to agents based on their ordinal prefer-
ences in a sequential manner. For each agent a
i
∈ A
(in a pre-defined order) the algorithm will allocate the
most preferred good g
(1)
of a
i
that is still available,
if g
(1)
is no longer available, allocate the next most
preferred good g
(2)
, and so on, until either a good is
allocated or all preferred goods are unavailable.
This greedy approach ensures that each agent re-
ceives the best possible good based on availability,
prioritizing agents in the order they are processed. In
order to keep the code simple, we used the command
Proceed to the next agent which tries to be equitable
enough and which can restart from the first agent if
there are unallocated goods and also agents a
i
with
O
i
̸= G. Obviously, this algorithm may not result in
optimal overall satisfaction.
4 PREDICTIVE-BASED
RESOURCE ALLOCATION
In this section, we introduce the concepts that under-
pin the two-step framework for refining multi-agent
resource allocation. The first step relies on the aggre-
gated preferences of agents, calculated using social
choice methods outlined in Section 4.1 and detailed
in (Brandt et al., 2016). The second step involves
the use of classifiers (Jordan and Mitchell, 2015), dis-
cussed in Section 4.2, to identify pairs of agents for
good-swapping, inspired by Condorcet principles.
Input: A = {a
1
,a
2
,. . . ,a
n
} (agents in a
predefined order),
G = {g
1
,g
2
,. . . , g
m
} (set of goods),
M = {m
1
,m
2
,. . . , m
m
} (multiplicities of goods),
V = {v
1
,v
2
,. . . , v
n
} (ordinal preference
functions for each agent).
Output: An allocation O = {O
1
,O
2
,. . . , O
n
}
where O
i
is the set of goods allocated
to agent a
i
.
Initialize O
i
←
/
0 for all a
i
∈ A ;
No goods ←
∑
m
j=1
m
j
;
while No goods > 0 do
foreach a
i
∈ A do
foreach g
j
∈ G in order given by v
i
do
if m
j
> 0 and g
j
/∈ O
i
then
O
i
← O
i
∪ {g
j
} ;
m
j
← m
j
− 1 ;
No goods ← No goods − 1 ;
Proceed to the next agent;
end
end
end
end
return O
Algorithm 1: Simple Allocation Algorithm.
4.1 Social Choice
Let us now define an aggregated preference v
A
that
combines the individual preferences v
1
,v
2
,. . . ,v
n
of
all agents A = {a
1
,a
2
,. . . ,a
n
} into a single collective
preference over the goods G = {g
1
,g
2
,. . . ,g
m
}. The
aggregated preference v
A
can be formally expressed
as:
v
A
= Aggregation(v
1
,v
2
,. . . ,v
n
),
where, Aggregation is the function (e.g., Borda or
Condorcet) used to combine the individual prefer-
ences into a collective ranking.
In the Borda method, a score is assigned to goods
based on their ordinal rankings in each agent’s prefer-
ICAART 2025 - 17th International Conference on Agents and Artificial Intelligence
658

ence. The aggregated preference v
A
is defined as:
v
A
(g
j
) =
n
∑
i=1
(n − v
i
(g
j
) + 1),
• v
i
(g
j
) is the rank of g
j
in agent a
i
’s preference list
(lower ranks indicate higher preference),
• (n − v
i
(g
j
) + 1) converts the rank into a score
(higher scores indicating higher preference).
The goods are then ordered in descending order of
v
A
(g
j
) to form the aggregated preference.
In the Condorcet method, the aggregated pref-
erence is determined through pairwise comparisons.
For each pair of goods (g
j
,g
k
, a preference matrix M
is constructed:
M(g
j
,g
k
) =
n
∑
i=1
1
v
i
(g
j
)<v
i
(g
k
)
,
• M(g
j
,g
k
) is the number of agents preferring g
j
to
g
k
,
• 1
v
i
(g
j
)<v
i
(g
k
)
= 1 if g
j
is ranked higher than g
k
by
agent a
i
, and 0 otherwise.
A good g
j
is a Condorcet winner if it is preferred to
every other good in pairwise comparisons:
M(g
j
,g
k
) > M(g
k
,g
j
) ∀g
k
̸= g
j
.
An aggregation function for collective prefer-
ences should adhere to several key principles to en-
sure fairness and rationality: Pareto efficiency, Non-
Dictatorship, Independence of Irrelevant Alternatives
(IIA), and Anonymity.
Pareto Efficiency requires that if all agents prefer
a good g
j
over another good g
k
, the collective prefer-
ence v
A
ranks g
j
higher than g
k
:
v
i
(g
j
) < v
i
(g
k
) ∀i ∈ A,
then the aggregated preference must satisfy:
v
A
(g
j
) < v
A
(g
k
).
Non-Dictatorship ensures that no single agent a
i
∈
A can unilaterally determine the collective preference
v
A
, unless their preferences align with all agents unan-
imous preference:
∃g
j
,g
k
∈ G such that v
i
(g
j
) < v
i
(g
k
) and
v
A
(g
k
) < v
A
(g
j
),
for at least one pair of goods g
j
,g
k
and a
i
.
Independence of Irrelevant Alternatives (IIA) en-
sures that the collective ranking of two goods g
j
and
g
k
depends on their relative rankings in individual
preferences, unaffected by the presence or absence of
other goods. Formally, if:
v
i
(g
j
) < v
i
(g
k
) ∀i ∈ A,
then the aggregated preference must also preserve this
ordering:
v
A
(g
j
) < v
A
(g
k
).
Anonymity requires that the aggregation mechanism
treats all agents equally, i.e. the outcome is invariant
to the permutation of agent indices.
The Condorcet method satisfies Pareto efficiency, as
any unanimously preferred good will dominate oth-
ers in pairwise comparisons. It also respects Non-
Dictatorship, as no single agent can dictate out-
comes unless their preferences align with the unan-
imous preference of all agents. Condorcet also sat-
isfies Anonymity, as all agents are treated symmet-
rically in the aggregation process. However, Con-
dorcet violates IIA because the addition or removal
of other goods can alter pairwise comparison results.
A Condorcet winner might not always exist (Arrow,
2012). The Borda method equally violates IIA and re-
spects Anonymity, as the scoring mechanism treats all
agents equally. However, Borda fails to satisfy Pareto
efficiency because the collective ranking can assign
higher aggregate scores to a good g
k
that is unani-
mously less preferred than g
j
. Furthermore, Borda
does not adhere to the Condorcet criterion, as it can
select a good that loses in pairwise comparisons to
another (Arrow, 2012).
As an example consider four agents A =
{a
1
,a
2
,a
3
,a
4
} and three goods G = {g
1
,g
2
,g
3
} with
the following preferences:
a
1
: g
1
≻ g
2
≻ g
3
, a
2
: g
2
≻ g
3
≻ g
1
,
a
3
: g
3
≻ g
1
≻ g
2
, a
4
: g
3
≻ g
2
≻ g
1
.
Using Borda points are assigned as 2 for first
place, 1 for second, and 0 for third:
g
1
: 3, g
2
: 4, g
3
: 5.
The Borda ranking is:
g
3
≻ g
2
≻ g
1
.
For Condorcet each pair of goods is compared
across agents:
g
1
vs. g
2
: 2 votes for each (Tie),
g
1
vs. g
3
: 2 votes for each (Tie),
g
2
vs. g
3
: 2 votes for each (Tie).
Since no good wins all pairwise comparisons, no
Condorcet winner exists.
4.2 Preference Classifiers
Classifiers in Machine Learning analyse data fea-
tures to learn patterns that distinguish between differ-
ent categories. Features are measurable attributes or
Prediction-Based Selective Negotiation for Refining Multi-Agent Resource Allocation
659

properties of the data that are considered relevant for
the classification task. The classifier evaluates these
features to identify patterns or decision boundaries
that separate classes.
A classifier can be formalized as a function f :
X → Y , where:
• X is the input space, representing the set of possi-
ble feature vectors, x = (x
1
,x
2
,. . . ,x
d
) ∈ X , where
d is the number of features.
• Y is the output space, which contains the set
of possible labels or classes, typically Y =
{1,2,. . . ,C}, where C is the number of classes.
• f (x;θ) is the decision function, parameterized by
θ, which maps input features x to a predicted class
ˆy ∈ Y .
The classifier is trained on a dataset D =
{(x
i
,y
i
)}
N
i=1
, where x
i
∈ X are feature vectors and
y
i
∈ Y are the corresponding ground truth labels. The
objective during training is to optimize the parameters
θ by minimizing a loss function L, i.e. the discrepancy
between the predicted labels ˆy
i
= f (x
i
;θ) and the true
labels y
i
:
ˆ
θ = argmin
θ
1
N
N
∑
i=1
L( f (x
i
;θ), y
i
).
The classifier can also incorporate a hypothesis
space H , representing the set of all possible decision
functions f that can be chosen given the parameteri-
zation:
f ∈ H , H = { f (x; θ) | θ ∈ Θ},
where, Θ is the parameter space.
Once trained, the classifier predicts the label for a
new input x
∗
by computing:
ˆy = argmax
y∈Y
P(y | x
∗
;
ˆ
θ),
where, P(y | x
∗
;
ˆ
θ) represents the model’s estimated
probability of class y for the input x
∗
, given the opti-
mized parameters
ˆ
θ.
The Predictive-Based Resource Allocation Problem
extends the traditional resource allocation problem by
incorporating classifiers to predict preferences based
on agent feature as follows:
(A, G, M,F , C ),
where:
• A = {a
1
,a
2
,. . . ,a
n
} is the set of n agents.
• G = {g
1
,g
2
,. . . ,g
m
} is the set of m goods.
• M = {m
1
,m
2
,. . . ,m
m
} specifies the maximum
number of agents m
j
that can receive each good
g
j
, satisfying
∑
m
j=1
m
j
≥ n.
• F is the feature space, where each agent a
i
is as-
sociated with a feature vector x
i
= ( f
i1
,. . . , f
ip
).
• C is the set of classifiers, where each classifier
f
c
k
: X → Y predicts the preference rankings of
agents based on their features.
For each agent a
i
, the predicted preference rank-
ing over goods is given by:
ˆv
i
= f
c
k
(x
i
),
where f
c
k
is a classifier applied to x
i
, and ˆv
i
defines
the predicted strict ranking of goods G.
An allocation O = {O
1
,. . . ,O
n
} maps subsets of
goods to agents, satisfying the multiplicity:
∑
a
i
∈A
1
g
j
∈O
i
≤ m
j
, ∀g
j
∈ G,
where 1
g
j
∈O
i
= 1 if g
j
∈ O
i
, and 0 otherwise.
The satisfaction of agent a
i
is a function S
i
that as-
sociates to each subset G
′
⊆ G of goods a positive real
value S
i
(G
′
). Note that here, the individual satisfac-
tion S
i
of agent a
i
depends on the predicted preference
ranking ˆv
i
. The satisfaction of an allocation O is the
sum of the satisfaction of the agents for their respec-
tive allocations: S(O) =
∑
n
i=1
S
i
(O
i
). The objective is
to find an allocation O
∗
that maximizes satisfaction.
5 OPTIMISING THE
ALLOCATION
The algorithm introduced in this section operates in
four phases, each presented as a distinct algorithm
for lisibility. The first phase, which corresponds to
the naive allocation based on the lexicographical or-
der of agents, is identical to the algorithm described
in Section 3 and is repeated here for readability. In
the second phase, the Borda aggregation is com-
puted, which is then used in the third phase to per-
form a Borda-based allocation of goods. Finally, in
the fourth phase, the allocation is refined through
classifier-driven swaps to further enhance satisfaction.
The last two phases form the two-step framework for
prediction-based refinement of the allocation.
In the second phase, the Borda method is applied
to aggregate individual preferences into a collective
ranking of goods. Each good receives a score based
on its rank in the agents’ preference lists. The goods
are then ordered by their scores to produce a collective
preference ranking that reflects the priorities of the
group.
ICAART 2025 - 17th International Conference on Agents and Artificial Intelligence
660

Input: A = {a
1
,a
2
,. . . , a
n
}: Set of agents;
G = {g
1
,g
2
,. . . , g
m
}: Set of goods;
M = {m
1
,m
2
,. . . , m
m
}: Multiplicities of goods;
{v
1
,v
2
,. . . , v
n
}: Agents’ preferences over
goods;
A lexicographical order on A;
Output: Initial allocation O = {O
1
,O
2
,. . . , O
n
}
Construct allocation O using Algorithm in
Section 3 or using any else Heuristic;
Compute overall satisfaction S(O) of the initial
allocation;
return O;
Algorithm 2: Phase 1: Initial Allocation.
Input: A = {a
1
,a
2
,. . . , a
n
}: Set of agents;
G = {g
1
,g
2
,. . . , g
m
}: Set of goods;
{v
1
,v
2
,. . . , v
n
}: Agents’ preferences over
goods;
Output: Aggregated preference v
A
foreach g
j
∈ G do
Compute Borda score:
B(g
j
) =
n
∑
i=1
score(g
j
in v
i
)
end
Sort G in descending order of Borda scores to
form aggregated preference v
A
;
return v
A
;
Algorithm 3: Phase 2: Aggregated Preferences (Borda
Method).
In the third phase, the algorithm refines the Borda
allocation to improve overall satisfaction according to
the collective preferences. Starting with the most pre-
ferred good in the collective ranking, the algorithm
allocates goods to agents in a way that maximizes the
number of agents receiving goods they rank highly.
This process continues for successive goods until ei-
ther the goods are exhausted or further allocations no
longer align with the collective ranking.
In the final phase, predictive models are used to
adjust the allocation further. For each agent, a classi-
fier predicts preferences based on their feature profile.
Agents whose predicted preferences differ from their
true preferences are identified, and the algorithm pro-
poses swaps with other agents in order to increase the
overall satisfaction of the allocation. The swaps are
done iteratively to improve the alignment of the allo-
cation. More precisely, a swap between agents a
i
and
a
k
means that goods ginO
i
and g
′
∈ O
k
are identified
such that
S
i
((O
i
− {g}) ∪ {g
′
}) + S
k
((O
k
− {g
′
}) ∪ {g}) >
S
i
(O
i
) + S
k
(O
k
).
Input: A = {a
1
,a
2
,. . . ,a
n
}: Set of agents;
G = {g
1
,g
2
,. . . ,g
m
}: Set of goods;
Aggregated preference v
A
;
Current allocation O;
{m
′
1
,m
′
2
,. . . ,m
′
m
}: Current multiplicity of
available (m
′
j
> 0) goods;
Output: Updated allocation O
foreach g
j
∈ G (in v
A
-descending order) do
foreach a
i
∈ A do
if m
′
j
> 0 and g
j
/∈ O
i
then
Allocate g
j
to a
i
: O
i
← O
i
∪ {g
j
};
Update m
′
j
← m
′
j
− 1;
end
end
end
return O;
Algorithm 4: Phase 3: Satisfaction Aggregated Preference
Boosting.
Input: A = {a
1
,a
2
,. . . ,a
n
}: Set of agents;
G = {g
1
,g
2
,. . . ,g
m
}: Set of goods;
Feature space F ;
Classifiers C ; Current allocation O;
Output: Optimized allocation O
∗
foreach a
i
∈ A do
Predict preferences using classifier:
ˆv
i
= f
c
k
(x
i
)
if ˆv
i
̸= v
i
then
Identify agents a
k
such that a swap with
a
i
is possible;
Propose swaps between a
i
and a
k
to
improve satisfaction;
Update O
i
and O
k
if swaps are accepted;
end
end
return O
∗
= O;
Algorithm 5: Phase 4: Optimization.
The algorithm outputs an optimized allocation that
improves on the naive approach by incorporating both
collective preferences and predictions. In each of the
Phases 1, 3 and 4, an implicit assumption is made:
each (individual) satisfaction function S
i
of the agent
a
i
is monotone:
if G
1
⊆ G
2
⊆ G then S
i
(G
1
) ≤ S
i
(G
2
).
Then, it is not difficult to see that the following theo-
rem holds.
Theorem 1. Let O be the naive initial allocation and
O
∗
the final allocation produced by the Predictive-
Based Resource Allocation Algorithm. If all satisfac-
tion function S
i
of the agents are monotone, then
S(O
∗
) ≥ S(O),
where S(·) is the overall satisfaction of an allocation.
Prediction-Based Selective Negotiation for Refining Multi-Agent Resource Allocation
661

6 FUTURE WORK
In this paper, we proposed a predictive-based resource
allocation algorithm that combines machine learning
predictions with preference aggregation techniques to
optimize resource allocation in multi-agent systems.
• Experimental validation is necessary to assess the
framework performance in diverse practical set-
tings (cloud resource allocation, logistics, pub-
lic goods distribution etc.). This includes a for-
mal computational complexity analysis evaluating
scalability of the proposed algorithm as well as
execution time, memory usage, and performance
to assess the scalability of your approach while
varying distributions of preferences and goods.
• Apart from the theoretical study of the properties
such as envy-freeness or equitability another no-
tion that should be investigated is that of “stable”
allocation (the agents do not have any incentive to
further swap).
• When several agents are eligible for a swap based
on their profiles and predicted preferences, crite-
ria for choosing the most appropriate participants
need to be defined. This decision introduces po-
tential concerns regarding ethics and bias, particu-
larly when prioritizing agents could inadvertently
favor certain groups over others (Hurwicz, 1973).
• Furthermore, the notion of overall satisfaction
could be refined to incorporate subpopulation-
specific goals. For instance, rather than optimiz-
ing global satisfaction, the algorithm could priori-
tize improving the satisfaction of specific subpop-
ulations (e.g., AI-agents vs humans in hybrid soci-
eties) based on implicit or explicit norms (Aldew-
ereld et al., 2016). Similarly to as above, this issue
is directly related to the fairness and ethical con-
cerns of our approach.
REFERENCES
Aldewereld, H., Dignum, V., and Vasconcelos, W. W.
(2016). Group norms for multi-agent organisations.
ACM Transactions on Autonomous and Adaptive Sys-
tems (TAAS), 11(2):1–31.
Arrow, K. J. (2012). Social choice and individual values,
volume 12. Yale university press.
Brandt, F., Conitzer, V., Endriss, U., Lang, J., and Procac-
cia, A. D. (2016). Handbook of computational social
choice. Cambridge University Press.
Chevaleyre, Y., Dunne, P. E., Endriss, U., Lang, J.,
Lemaitre, M., Maudet, N., Padget, J., Phelps, S.,
Rodr
´
ıgues-Aguilar, J. A., and Sousa, P. (2005). Issues
in multiagent resource allocation.
Croitoru, M. and Croitoru, C. (2011). Generalised net-
work flows for combinatorial auctions. In 2011
IEEE/WIC/ACM International Conferences on Web
Intelligence and Intelligent Agent Technology, vol-
ume 2, pages 313–316. IEEE.
Dafoe, A., Hughes, E., Bachrach, Y., Collins, T., Mc-
Kee, K. R., Leibo, J. Z., Larson, K., and Graepel,
T. (2020). Open problems in cooperative ai. arXiv
preprint arXiv:2012.08630.
Hurwicz, L. (1973). The design of mechanisms for resource
allocation. The American Economic Review, 63(2):1–
30.
Ibaraki, T. and Katoh, N. (1988). Resource allocation prob-
lems: algorithmic approaches. MIT press.
Jordan, M. I. and Mitchell, T. M. (2015). Machine learn-
ing: Trends, perspectives, and prospects. Science,
349(6245):255–260.
Katoh, N. and Ibaraki, T. (1998). Resource allocation
problems. Handbook of Combinatorial Optimization:
Volume1–3, pages 905–1006.
Li, J., Ni, X., Yuan, Y., and Wang, F.-Y. (2018). A hier-
archical framework for ad inventory allocation in pro-
grammatic advertising markets. Electronic Commerce
Research and Applications, 31:40–51.
Vinothina, V. V., Sridaran, R., and Ganapathi, P. (2012). A
survey on resource allocation strategies in cloud com-
puting. International Journal of Advanced Computer
Science and Applications, 3(6).
Zhang, J., Zhang, M., Ren, F., and Liu, J. (2016). An in-
novation approach for optimal resource allocation in
emergency management. IEEE Transactions on Com-
puters.
ICAART 2025 - 17th International Conference on Agents and Artificial Intelligence
662