New Zealand

James M. Hereford and Michael A. Siebold (March 1st 2010). Bio-Inspired Search Strategies for Robot Swarms, Swarm Robotics from Biology to Robotics, Ester Martinez Martin, IntechOpen, DOI: 10.5772/8600. 

Andrej Gams, Tadej Petrič, Aleš Ude and Leon Žlajpah (January 20th 2012). Performing Periodic Tasks: On-Line Learning, Adaptation and Synchronization with External Signals, The Future of Humanoid Robots - Research and Applications, Riadh Zaier, IntechOpen, DOI: 10.5772/26544. 

Xavier Giralt, Albert Hernansanz, Alberto Rodriguez and Josep Amat (October 1st 2008). Robotic Proximity Queries Library for Online Motion Planning Applications, New Developments in Robotics Automation and Control, Aleksandar Lazinica, IntechOpen, DOI: 10.5772/6287. 

Swarms can be applied in many relevant domains, such as patrolling or rescue. They usually follow simple local rules, leading to complex emergent behavior. Given their wide applicability, an agent may need to take decisions in an environment containing a swarm that is not under its control, and that may even be an antagonist. Predicting the behavior of each swarm member is a great challenge, and must be done under real time constraints, since they usually move constantly following quick reactive algorithms. We propose the first two solutions for this novel problem, showing integrated on-line learning and planning for decision-making with unknown swarms: (i) we learn an ellipse abstraction of the swarm based on statistical models, and predict its future parameters using time-series; (ii) we learn algorithm parameters followed by each swarm member, in order to directly simulate them. We find in our experiments that we are significantly faster to reach an objective than local repulsive forces, at the cost of success rate in some situations. Additionally, we show that this is a challenging problem for reinforcement learning.


AAMAS 2020

Real-time Learning and Planning in Environments with Swarms: A Hierarchical and a Parameter-based Simulation Approach

on-line planning

on-line learning

swarms

It is my great pleasure and honor to welcome you to the 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026), held in San Diego, California, USA, July 2–7, 2026. ACL 2026 continues our fields tradition of excellence in scholarship, innovation, and inclusivity, and I am deeply grateful to the many volunteers who have worked tirelessly to bring this event to life.

It is exciting to see the many developments in the field reflected at our conference, such as the advancement of large language models (born out of our community) into many application directions. This is not without challenges, partly due to unsolved problems, and partly due to misuse of the technology. All this is reflected well at ACL 2026. The conference also reflects that the field is not only about advancing technology, but also the science of fundamentally understanding language and how it is use by many different speaker communities. <br>
[Full Message from General Chair Philipp Koehn](https://drive.google.com/file/d/1QiJFOmMjvDgsDo3i1XzlQnImxe3RU_Cy/view?usp=sharing) <br><br><br>
<html><button style="display: inline-flex; align-items: center; justify-content: center; white-space: nowrap; border-radius: 9999px; font-weight: bold; background: #7c3aed; color: white; font-family: 'Space Grotesk', sans-serif; height: 40px; font-size: 16px; padding: 0 20px; border: none; cursor: pointer" onclick="window.open('https://drive.google.com/file/d/1QiJFOmMjvDgsDo3i1XzlQnImxe3RU_Cy/view?usp=sharing')">Conference Handbook</button></html>
<br><br><br><br>

To access this event page, you need to log in with the **email address you registered with**. <br>Access credentials will be sent to your email from Underline -  subject line "Welcome to ACL 2026". Please be sure to check your spam email folder if you do not see an email confirmation right away.

To access this event page, you are required to register.
Please complete your registration to continue.

We recommend reading [**the registration information**](https://2026.aclweb.org/registration/) first.

**Online Registration Form**: https://acl.swoogo.com/acl2026

Registration Required

ACL 2026

The 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026) will take place in San Diego, California, from July 2 - 7, 2026.

technical paper

AAMAS is the leading scientific conference for research in autonomous agents and multi-agent systems. The AAMAS conference series was initiated in 2002 as the merging of three respected scientific meetings: the International Conference on Multi-Agent Systems (ICMAS), the International Workshop on Agent Theories, Architectures, and Languages (ATAL), and the International Conference on Autonomous Agents (AA). The aim of the joint conference is to provide a single, high-profile, internationally-respected archival forum for scientific research in the theory and practice of autonomous agents and multi-agent systems.

Browse keynotes, discussions, panels and over 300 presentations.


AAMAS is the leading scientific conference for research in autonomous agents and multi-agent systems. The AAMAS conference series was initiated in 2002 as the merging of three respected scientific meetings: the International Conference on Multi-Agent Systems (ICMAS), the International Workshop on Agent Theories, Architectures, and Languages (ATAL), and the International Conference on Autonomous Agents (AA). 

This paper investigates how to efficiently transition and update policies, trained initially with demonstrations, using off-policy actor-critic reinforcement learning. It is well-known that techniques based on Learning from Demonstrations, for example behavior cloning, can lead to proficient policies given limited data. However, it is currently unclear how to efficiently update that policy using reinforcement learning as these approaches are inherently optimizing different objective functions. Previous works have used loss functions, which combine behavior cloning losses with reinforcement learning losses to enable this update. However, the components of these loss functions are often set anecdotally, and their individual contributions are not well understood. In this work, we propose the Cycle-of-Learning (CoL) framework that uses an actor-critic architecture with a loss function that combines behavior cloning and 1-step Q-learning losses with an off-policy pre-training step from human demonstrations. This enables transition from behavior cloning to reinforcement learning without performance degradation and improves reinforcement learning in terms of overall performance and training time. Additionally, we carefully study the composition of these combined losses and their impact on overall policy learning. We show that our approach outperforms state-of-the-art techniques for combining behavior cloning and reinforcement learning for both dense and sparse reward scenarios. Our results also suggest that directly including the behavior cloning loss on demonstration data helps to ensure stable learning and ground future policy updates.


Integrating Behavior Cloning and Reinforcement Learning for Improved Performance in Dense and Sparse Reward Environments

We present a general model for the operation of a cloud computing server comprised of one or more speed-scalable processors. Typically, tasks are submitted to such a cloud computing server in an online fashion, and the server operator has to schedule the tasks and decides on payments without knowledge about the tasks arriving in the future. Although very natural, this cloud computing problem on speed-scalable processors has not been studied from a mechanism design perspective in the online setting.

We provide a mechanism for this setting, both for a single and multiprocessor environment, that has several desirable properties: (1) the induced game admits a subgame perfect equilibrium in pure strategies and therefore a pure Nash equilibrium, (2) the Price of Anarchy is constant, (3) the mechanism is budget balanced, i.e., the sum of the payments of the agents is equal to the total energy costs, (4) the communication complexity is low, (5) the mechanism is computationally tractable for both the service operator and the agents, and (6) the agents' payment is also intuitive and easy to communicate to them. We also provide a second mechanism with a better Price of Anarchy, which in turn is more involved to implement.

We are able to extend our mechanisms and results to the Bayesian setting, where the type of each agent is drawn independently from some underlying distribution and agents are minimizing their expected costs. In this setting we also show the same approximation factor of our mechanism as in the basic online setting in both the single and the multiprocessor environment.


A General Framework for Energy-Efficient Cloud Computing Mechanisms

Moving nodes in a Mobile Wireless Sensor Network (MWSN) typically have two maintenance objectives: (i) extend the coverage of the network as long as possible to a target area, and (ii) extend the longevity of the network as much as possible. 
As nodes move and also route traffic in the network, their battery levels deplete differently for each node. 
Dead nodes lead to loss of connectivity and even to disengaging full parts of the network. 
Several reactive and rule-based approaches have been proposed to solve this issue by adapting redeployment to depleted nodes. 
However, in large networks a cooperative approach may increase performance by taking the evolution of node battery and traffic into account. 
In this paper, we present a hybrid agent-based architecture that addresses the problem of depleting nodes during the maintenance phase of a MWSN. 
Agents, each assigned to a node, collaborate and adapt their behaviour to their battery levels. 
The collaborative behavior is modeled through the willingness to interact abstraction, which defines when agents ask and give help to one another. 
Thus, depleting nodes may ask to be replaced by healthier counterparts and move to areas with less traffic or to a collection point. 
At the lower level, negotiations trigger a reactive navigation behaviour based on Social Potential Fields (SPF). 
It is shown that the proposed method improves coverage and extends network longevity in an environment without obstacles as compared to SPF alone.

Adaptive Autonomy in Wireless Sensor Networks

Reinforcement learning has been successful in training autonomous agents to accomplish goals in complex environments. Although this has been adapted to multiple settings, including robotics and computer games, human players often find it easier to obtain higher rewards in some environments than reinforcement learning algorithms. This is especially true of high-dimensional state spaces where the reward obtained by the agent is sparse or extremely delayed. In this presentation, we introduce the FRESH (Feedback-based REward SHaping) framework, which eﬀectively integrates feedback signals supplied by a human operator with deep reinforcement learning algorithms in high-dimensional state spaces. During training, a human operator is presented with trajectories from a replay buﬀer and then provides feedback on states and actions in the trajectory. In order to generalize feedback signals provided by the human operator to previously unseen states and actions at test-time, we use a feedback neural network. We use an ensemble of neural networks with a shared network architecture to represent model uncertainty and the confidence of the neural network in its output. The output of the feedback neural network is converted to a shaping reward that is augmented to the reward provided by the environment. We evaluate our approach on the Bowling and Skiing Atari games in the arcade learning environment. Although human experts have achieved high scores in these environments, state-of-the-art deep learning algorithms perform poorly. We observe that FRESH achieves much higher scores than state-of-the-art deep learning algorithms in both environments. FRESH also achieves a 21.4% higher score than a human expert in Bowling and does as well as an expert in Skiing.

FRESH: Interactive Reward Shaping in High-Dimensional State Spaces using Human Feedback

When multiple agents learn in a decentralized manner, the environment appears non-stationary from the perspective of an individual agent due to the exploration and learning of the other agents. Recently proposed deep multi-agent reinforcement learning methods have tried to mitigate this non-stationarity by attempting to determine which samples are from other agent exploration or suboptimality and take them less into account during learning. Based on the same philosophy, this paper introduces a decentralized quantile estimator, which aims to improve performance by distinguishing non-stationary samples based on the likelihood of returns. In particular, each agent considers the likelihood that other agent exploration and policy changes are occurring, essentially utilizing the agent’s own estimations to weigh the learning rate that should be applied towards the given samples. We introduce a formal method of calculating differences of our return distribution representations and methods for utilizing it to guide updates. We also explore the effect of risk-seeking strategies for adjusting learning over time and propose adaptive risk distortion functions which guides risk sensitivity. Our experiments, on traditional benchmarks and new domains, show our methods are more stable, sample efficient and more likely to converge to a joint optimal policy than previous methods.


Likelihood Quantile Networks for Coordinating Multi-Agent Reinforcement Learning

Previous approaches to adversary modeling in network security games (NSGs) have been caught in the paradigm of first building a full adversary model, either from expert input or historical attack data, and then solving the game. Motivated by the need to disrupt the multibillion dollar illegal smuggling networks, such as wildlife and drug trafficking, this paper introduces a fundamental shift in learning adversary behavior in NSGs by focusing on the accuracy of the model using the downstream game that will be solved. Further, the paper addresses technical challenges in building such a game-focused learning model by i) applying graph convolutional networks to NSGs to achieve tractability and differentiability and ii) using randomized block updates of the coefficients of the defender's optimization in order to scale the approach to large networks. We show that our game-focused approach yields scalability and higher defender expected utility than models trained for accuracy only.

Scalable Game-Focused Learning of Adversary Models: Data-to-Decisions in Network Security Games

A Qualitative Approach to Composing Value-Align Norm Systems

State-of-the-art multi-agent reinforcement learning has achieved remarkable success in recent years. The success has been mainly based on the assumption that all teammates perfectly cooperate to optimize a global objective in order to achieve a common goal. While this may be true in the ideal case, these approaches could fail in practice, since in multi-agent systems (MAS), all agents may be a potential source of failure. In this presentation, we focus on resilience in cooperative MAS and propose an Antagonist-Ratio Training Scheme (ARTS) by reformulating the original target MAS as a mixed cooperative-competitive game between a group of protagonists which represent agents of the target MAS and a group of antagonists which represent failures in the MAS. While the protagonists can learn robust policies to ensure resilience against failures, the antagonists can learn malicious behavior to provide an adequate test suite for other MAS. We empirically evaluate ARTS in a cyber physical production domain and show the effectiveness of ARTS w.r.t. resilience and testing capabilities.


Learning and Testing Resilience in Cooperative Multi-Agent Systems

Multi-agent resource allocation is an important and well-studied problem within AI and economics. It is generally assumed that the quantity of each resource is known a priori. However, in many real-world problems, such as the production of renewable energy which is typically weather dependent, the exact amount of each resource may not be known at the time of decision making. In this paper we investigate fair division of a homogeneous divisible resource where the available amount is given by a probability distribution.
Specifically, we study the notion of ex-ante envy-freeness, where, in expectation, agents weakly prefer their allocation over every other agent's allocation. We analyse the trade-off between fairness and social welfare. We show that allocations satisfying ex-ante envy-freeness can result in higher social welfare compared to those satisfying ex-post envy-freeness. Nevertheless, the price of envy-freeness is at least $\Omega(n)$, where $n$ is the number of agents, and this is tight under concave valuation functions. Principally, we show that the problem of optimising ex-ante social welfare subject to ex-ante envy-freeness is NP-hard in the strong sense. Finally, we devise an integer program to calculate the optimal ex-ante envy-free allocation for linear satiable valuation functions.


Fair Allocation of Resources with Uncertain Availability

As part of the DARPA SocialSim challenge, we address the problem of predicting behavioral phenomena including information spread involving hundreds of thousands of users across three major linked social networks: Twitter, Reddit and GitHub. Our approach develops a framework for data-driven agent simulation that begins with a discrete-event simulation of the environment populated with generic, ﬂexible agents, then optimizes the decision model of the agents by combining a number of machine learning classification problems. The ML problems predict when an agent will take a certain action in its world and are designed to combine aspects of the agents, gathered from historical data, with dynamic aspects of the environment including the resources, such as tweets, that agents interact with at a given point in time. In this way, each of the agents makes individualized decisions based on their environment, neighbors and history during the simulation, although global simulation data is used to learn accurate generalizations. This approach showed the best performance of all participants in the DARPA challenge across a broad range of metrics. We describe the performance of models both with and without machine learning on measures of cross-platform information spread defned both at the level of the whole population and at the community level. The best-performing model overall combines learned agent behaviors with explicit modeling of bursts in global activity. Because of the general nature of our approach, it is applicable to a range of prediction problems that require modeling individualized, situational agent behavior from trace data that combines many agents.


Downloads

Next from AAMAS 2020

Integrating Behavior Cloning and Reinforcement Learning for Improved Performance in Dense and Sparse Reward Environments

Similar lecture

Data Swarms: Optimizable Generation of Synthetic Evaluation Data

.css-70qvj9{display:-webkit-box;display:-webkit-flex;display:-ms-flexbox;display:flex;-webkit-align-items:center;-webkit-box-align:center;-ms-flex-align:center;align-items:center;}Downloads

Next from AAMAS 2020

Integrating Behavior Cloning and Reinforcement Learning for Improved Performance in Dense and Sparse Reward Environments

Similar lecture

Data Swarms: Optimizable Generation of Synthetic Evaluation Data

Downloads