New Zealand

Khair Eddin Sabri, Ridha Khedri and Jason Jaskolka (October 1st 2009). Algebraic Model for Agent Explicit Knowledge in Multi-agent Systems, Advanced Technologies, Kankesu Jayanthakumaran, IntechOpen, DOI: 10.5772/8211. 

Ilaria Giannoccaro and Pierpaolo Pontrandolfo (February 1st 2008). How Negotiation Influences the Effective Adoption of the Revenue Sharing Contract: A Multi-Agent Systems Approach, Supply Chain, Vedran Kordic, IntechOpen, DOI: 10.5772/5337.

Dennis Barrios-Aranibar and Luiz M. G. Gon&#231;alves (January 1st 2009). Influence Value Q-Learning: A Reinforcement Learning Algorithm for Multi Agent Systems, Theory and Novel Applications of Machine Learning, Meng Joo Er and Yi Zhou, IntechOpen, DOI: 10.5772/6675. 

We propose a deep reinforcement learning algorithm for semi-cooperative multi-agent tasks, where agents are equipped with their separate reward functions, yet with some willingness to cooperate. It is intuitive that defining and directly maximizing a global reward function leads to cooperation because there is no concept of selfishness among agents. However, it may not be the best way of inducing such cooperation due to problems that arise from training multiple agents with a single reward (e.g., credit assignment). In addition, agents may intentionally be given separate reward functions to induce task prioritization whereas a global reward function may be difficult to define without diluting the effect of different tasks and causing their reward factors to be disregarded. Our algorithm, called Peer Evaluation-based Dual DQN (PED-DQN), proposes to give peer evaluation signals to observed agents, which quantify how they strategically value a certain transition. This exchange of peer evaluation among agents over time turns out to render agents to gradually reshape their reward functions so that their action choices from the myopic best response tend to result in a more cooperative joint action.

AAMAS 2020

Inducing Cooperation through Reward Reshaping based on Peer Evaluations in Deep Multi-Agent Reinforcement Learning

cooperation

multi-agent reinforcement learning

multi-agent systems

reinforcement learning

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). 

technical paper

Given the preferences of several agents over a set of alternatives, there may be competing views on which of the alternative would be “best” to choose. We propose a formal model, grounded in social choice theory, for providing a justification for a given choice in the context of a given corpus of basic normative principles (so-called axioms) on which to base any possible step-by-step explanation for why a given target alternative should be elected in a given situation. Thus, our notion of justification has both an explanatory and a normative component. We also develop an algorithm for computing such justifications that exploits the analogy between
the notion of explanation and the concept of minimal unsatisfiable subset used in constraint programming. Finally, we report on an application of a proof-of-concept implementation of our approach for an experimental study of the explanatory power of several axioms proposed in the social choice literature.


Automated Justification of Collective Decisions via Constraint Solving

In this work, we study the fair resource sharing problem, where a set of resources needs to be shared by a set of agents. Each agent is unit-demand and each resource can serve a limited number of agents. The agents have (heterogeneous) preferences for the resources, and preferences for other agents with whom they share the resources. Our definition of fairness is mainly captured by envy-freeness. Due to the fact that an envy-free assignment may not exist even in simple settings, we propose a way to relax the definition: Pareto envy-freeness, where an assignment is Pareto envy-free if for any two agents i and j, agent i does not envy agent j for her received resource or the set of agents she shares the resource with. We study
to what extent Pareto envy-free assignments exist. Particularly, we are interested in a typical model, dorm assignment problem, where a number of students need to be accommodated to the dorms with the same capacity and the students’ preferences for dorm-mates are binary. We show that when the capacities of the dorms are 2, a Pareto envy-free assignment always exists and can be found in polynomial time; however, if the capacities increase to 3, Pareto envy-freeness cannot be guaranteed any more.


Fair Resource Sharing and Dorm Assignment

Counterfactual Regret Minimization (CFR) has found success in settings like poker which have both terminal states and perfect recall. We seek to understand how to relax these requirements. As a first step, we introduce a simple algorithm, local no-regret learning (LONR), which uses a Q-learning-like update rule to allow learning without terminal states or perfect recall. We discuss its provable convergence for the basic case of MDPs and present empirical results showing that it achieves last iterate convergence in NoSDE games, a class of general sum Markov games specifically designed to be challenging to learn where no prior algorithm is known to achieve convergence to a stationary equilibrium even on average. 


Combining No-Regret and Q-Learning

This presentation is about our AAMAS 2020 paper ""Plannable Approximations to MDP Homomorphisms: Equivariance under Actions"". This work exploits action equivariance for representation learning in reinforcement learning. Equivariance under actions states that
transitions in the input space are mirrored by equivalent transitions in latent space, while the map and transition functions should also commute. We introduce a contrastive loss function that enforces action equivariance on the learned representations. We prove that when our loss is zero, we have a homomorphism of a deterministic Markov Decision Process (MDP). Learning equivariant maps leads to structured latent spaces, allowing us to build a model on which we plan through value iteration. We show experimentally that for deterministic MDPs, the optimal policy in the abstract MDP can be successfully lifted to the original MDP. Moreover, the approach easily adapts to changes in the goal states. Empirically, we show that in such MDPs, we obtain better representations in fewer epochs compared to representation learning approaches using reconstructions, while generalizing better to new goals than model-free approaches.


Plannable Approximations to MDP Homomorphisms: Equivariance under Actions

Consider a closed environment with static obstacles and mobile agents moving around. There are hider agents that hide from the seeker agents. The seeker has a limited visibility range, and if a hider comes into the visibility region of a seeker, it is considered caught. The practical applications range from gaming to security. In this work, we focus on deterministic capture of hiders, even if they are guided by an Oracle which knows the future positions of seekers. We develop strategies for seekers, having limited visibility ranges, to catch all hiders and establish minimum bounds on the number of seekers required to catch the hiders, on a per strategy basis. We use spatio-temporal graph models and reasoning to formulate and address the problem.


Capturing Oracle Guided Hiders

Effective coordination is a critical requirement for human teaming, and is increasingly needed in teams of humans and robots. Building on decades of work in the behavioral literature, we have implemented a computational framework for coordination based on Shared Mental Models (SMMs) in which robots use a distributed knowledge base to coordinate activity. We also built a novel system connecting the robotic architecture, DIARC, to the 3D simulation environment, Unity, to serve as an evaluation platform for the framework implementation, and also for more general explorations of teaming with autonomous robots. Using this platform, we ran a user study to evaluate the framework by comparing performance of teams in which the robots used SMMs with those that did not. We found that teams in which the robots used SMMs significantly outperformed those without SMMs. This represents the first empirical demonstration that SMMs can be successfully used by fully autonomous robots interacting in natural language to improve team performance, bringing robots a step closer to genuine teammate


Toward Genuine Robot Teammates: Improving Human-Robot Team Performance Using Robot Shared Mental Models

An important aspect of multi-agent systems concerns the formation of coalitions that are stable or optimal in some well-defined way. The notion of popularity has recently received a lot of attention in this context. A partition is popular if there is no other partition in which more agents are better off than worse off. 
In 2019, a long-standing open problem concerning popularity was solved by proving that computing popular partitions in roommate games is NP-hard, even when preferences are strict. We show that this result breaks down when allowing for randomization: mixed popular partitions can be found efficiently via linear programming and a separation oracle. Mixed popular partitions are particularly attractive because they are guaranteed to exist in any coalition formation game.
Our result implies that one can efficiently verify whether a given partition in a roommate game is popular and that strongly popular partitions can be found in polynomial time (resolving an open problem). By contrast, we prove that both problems become computationally intractable when moving from coalitions of size 2 to coalitions of size 3, even when preferences are strict and globally ranked.


Finding and Recognizing Popular Coalition Structures

Designers of virtual agents have a combinatorically large space of choices for different media that comprise the look and behavior of their characters. We explore the systematic manipulation of animation quality, speech quality, and rendering style, and its impact on the perceptions of virtual agents in terms of naturalness, engagement, trust, credibility, and persuasion within a health counseling domain. The agent’s counseling behavior was based on live video footage of a human counselor. We conducted a between-subjects study that had 12 conditions. Character animation was varied between a static image, procedural animation using a gestuary, and manually rotoscoped animation. Voice quality was varied between recorded audio of the human counselor and synthetic speech. Character rendering style was varied between 3D-shaded realistic and toon-shaded. Prior studies indicate that people prefer and attribute more sociality to other people and agents when modalities are consistent in their level of quality. Thus, we hypothesize that people will be most affected by agents whose animation, voice, and rendering style are consistent, rather than the effects of channel quality being purely additive. Results indicate that natural animations and recorded voice are more appropriate for general acceptance, trust, and credibility of the agent, and how appropriate she seems for the task. However, our results indicate that for a brief health counseling task, animation might actually be distracting from the persuasive message, with the highest levels of persuasion found when the amount of agent animation is minimized.


Navigating the Combinatorics of Virtual Agent Design Space to Maximize Persuasion

We initiate the study of cloning in multiwinner elections, focusing
on single-transferable vote (STV), single-nontransferable vote
(SNTV), bloc, k-Borda, t-approval-CC, and Borda-CC. Transferring
the model of cloning due to Elkind et al. [15] from single-winner to
multiwinner elections, we consider decision problems describing
possible and necessary cloning in the zero-cost, the unit-cost, and
the general-cost model and study their computational complexity.
We show that, depending on the multiwinner voting rule and on
the cost model chosen, some of these cloning problems are in P,
some are NP-hard, and some of the latter (for which, in fact, already
winner determination is NP-hard) are fixed-parameter tractable.

The Complexity of Cloning Candidates in Multiwinner Elections

The hypergraph matching game is a cooperative game defined on a hypergraph such that the vertices are the players, and the characteristic function is the value of a maximum hypergraph matching on a hypergraph induced by a coalition. 
In this talk, we study a computationally tractable condition of the hypergraph matching game, called the convexity.
First, we prove that the problem of checking whether a given hypergraph matching game is convex or not is solvable in polynomial time.
Second, we prove that the Shapley value of a given convex hypergraph matching game is exactly computable in polynomial time.
Finally, we consider the fractional hypergraph matching game and prove that if the fractional hypergraph matching game is convex, then its characteristic function coincides with the characteristic function of the corresponding (integral) hypergraph matching game.

Downloads

Next from AAMAS 2020

Automated Justification of Collective Decisions via Constraint Solving

Similar lecture

Gifting in Multi-Agent Reinforcement Learning

.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

Automated Justification of Collective Decisions via Constraint Solving

Similar lecture

Gifting in Multi-Agent Reinforcement Learning

Downloads