United Kingdom

In many risk-aware and multi-objective reinforcement learning settings, the utility of the user is derived from the single execution of a policy. In these settings, making decisions based on the average future returns is not suitable. For example, in a medical setting a patient may only have one opportunity to treat their illness. When making a decision, just the expected return -- known in reinforcement learning as the value -- cannot account for the potential range of adverse or positive outcomes a decision may have. Our key insight is that we should use the distribution over expected future returns differently to represent the critical information that the agent requires at decision time.

AAMAS 2021

Distributional Monte Carlo Tree Search for Risk-Aware and Multi-Objective Reinforcement Learning

AAMAS is the largest and most influential conference in the area of agents and multiagent systems, bringing together researchers and practitioners in all areas of agent technology and providing and internationally renowned high-profile forum for publishing and finding out about the latest developments in the field. AAMAS is the flagship conference of the non-profit International Foundation for Autonomous Agents and Multiagent Systems (IFAAMAS).  

![](https://assets.underline.io/uploads/markdown_image/1/image/ce4604bd54468a006164d16e080204fb.jpg)

Please click on the button below to register for this event.

If you are registered and see this page please check the account email you are logged in with. It must be the same as for the registration.

If you still have trouble accessing the contact please contact support at aamas2021@underline.io 

poster

For intelligent conversational agents to speak about albums of images with users as humans do, they must be able to make sense of images as humans do. Computer vision methods can report directly observable information, but human beings care about more than the directly observable; they value holistic narratives that include affective and motivational evaluations, casual connections, and other inferred relationships from external knowledge. Drawing from theories in cognitive sensemaking and narrative coherence, we propose an approach for image sequence understanding that strives to generate and evaluate hypotheses about the relationships between people, events, and objects in images using commonsense knowledge, which are formed into a consistent network of hypotheses and observed facts via multi-objective optimization. The result is an enriched knowledge representation in the form of a knowledge graph which may later be used by a conversational agent.

Image Sequence Understanding through Narrative Sensemaking

Autonomous agents that can engage in social interactions with a human is the ultimate goal of a myriad of applications. A key challenge in the design of these applications is to define the social behavior of the agent, which requires extensive content creation. In this research, we explore how we can leverage current state-of-the-art tools to make inferences about the emotional state of a character in a story as events unfold, in a coherent way. We propose a character role-labelling approach to emotion tracking that accounts for the semantics of emotions. We show that, by identifying actors and objects of events and considering the emotional state of the characters, we can achieve better performance in this task when compared to end-to-end approaches.

CHARET: Character-centered Approach to Emotion Tracking in Stories

We study fair and economically efficient allocation of indivisible goods among agents whose valuations are rank functions of matroids. Such valuations constitute a well-studied class of submodular functions (i.e., they exhibit a diminishing returns property) and model preferences in several resource-allocation settings. We prove that, for matroid-rank valuations, a social welfare-maximizing allocation that gives each agent her maximin share always exists. Furthermore, such an allocation can be computed in polynomial time. We establish similar existential and algorithmic results for the pairwise maximin share guarantee as well.

Existence and Computation of Maximin Fair Allocations Under Matroid-Rank Valuations

We study the problem of spending a budget that is common to n agents. Agents submit demands to a central planner who uses the budget to fund a subset of them. The utility of an agent is the part of the budget spent on her own accepted demands. In a fair solution, the successful demands of each agent would represent a 1/n fraction of the budget. However, this is rarely possible because every demand is indivisible, i.e. either accepted in its entirety or rejected. We are interested in worst-case bounds on the largest proportion of the budget that is dedicated to the least funded agent. Our approach is not to solve the corresponding max min problem for every instance, but to tackle the problem from a higher level. The size of the largest demand compared to the budget and the number of agents, are two parameters that significantly influence how much the worst-off agent gets. We propose worst-case bounds on the best utility of the least funded agent for the class of instances where the number of agents and the most expensive demand are fixed to given values. A characterization of this quantity is provided for 1 and 2 agents. For more than 2 agents, %our results we propose lower and upper bounds that constitute a 14/15-approximation of the optimal value. Every existence result is complemented with a polynomial algorithm that builds a feasible solution satisfying our bounds.

Worst-case Bounds for Spending a Common Budget

Recent advances in the blockchain research have been made in two important directions. One is refined resilience analysis utilizing game theory to study the consequences of selfish behavior of users (miners), and the other is the extension from a linear (chain) structure to a non-linear (graphical) structure for performance improvements, such as IOTA and Graphcoin. The first question that comes to mind is ? what improvements that a blockchain system would see by leveraging these new advances. In this paper, we consider three major properties for a blockchain system: ?-partial verification, scalability, and finality-duration. We establish a formal framework and prove that no blockchain system can achieve ?-partial verification for any fixed constant ?, high scalability, and low finality-duration simultaneously. We observe that classical blockchain systems like Bitcoin achieves full verification (?=1) and low finality-duration, Ethereum 2.0 Sharding achieves low finality-duration and high scalability. We are interested in whether it is possible to partially satisfy the three properties.

A Game Theoretical Analysis of Non-Linear Blockchain System

In this paper we present Probabilistic Control Argumentation Frameworks (PCAFs) that extend classical Control Argumentation Frameworks (CAFs) to take into account probabilistic information in the reasoning process. We show that probabilities can be used to optimally control CAFs that cannot be controlled otherwise. We introduce the notion of controlling power, that represents the probability that a control configuration reaches its target. A computational method based on Monte Carlo simulations for computing the controlling power of control configurations is defined. We experimentally show that PCAFs outperform w.r.t runtime classical CAFs and in a large number of situations they can reach the target with a high probability while the classical CAFs fail.

Probabilistic Control Argumentation Frameworks

Ranking the participants of a tournament has applications in voting, paired comparisons analysis, sports and other domains. In this paper we introduce bipartite tournaments, which model situations in which two different kinds of entity compete indirectly via matches against players of the opposite kind; examples include education (students/exam questions) and solo sports (golfers/courses). In particular, we look to find rankings via chain  graphs, which correspond to bipartite tournaments in which the sets of adversaries defeated by the players on one side are nested with respect to set inclusion. Tournaments of this form have a natural and appealing ranking associated with them. We apply chain editing -- finding the minimum number of edge changes required to form a chain graph -- as a new mechanism for  tournament ranking. The properties of these rankings are investigated in a probabilistic setting, where they arise as maximum likelihood estimators, and through the axiomatic method of social choice theory. Despite some nice  properties, two problems remain: an important anonymity axiom is violated, and chain editing is NP-hard. We address both issues by relaxing the minimisation constraint in chain editing, and characterise the resulting ranking methods via  a greedy approximation algorithm.

Rankings for Bipartite Tournaments via Chain Editing

In recent years, machine learning (ML) models have been successfully applied in a variety of real-world applications. However, they are often complex and incomprehensible to human users. This can decrease trust in their outputs and render their usage in critical settings ethically problematic. As a result, several methods for explaining such ML models have been proposed recently, in particular for black-box models such as deep neural networks (NNs). Nevertheless, these methods predominantly explain model outputs in terms of inputs, disregarding the inner workings of the ML model computing those outputs. We present Argflow, a toolkit enabling the generation of a variety of ‘deep’ argumentative explanations (DAXs) for outputs of NNs on classification tasks.

Argflow: A Toolkit for Deep Argumentative Explanations for Neural Networks

In this study, we leverage multi-objective reinforcement learning to create tunable agents, i.e., agents that can adopt a range of different behaviours according to the designer's preferences, without the need for retraining. We apply this technique to sequential social dilemmas, settings where there is inherent tension between individual and collective rationality. Learning a single fixed policy in such settings leaves one at a significant disadvantage if the opponents' strategies change after learning is complete. In our work, we demonstrate empirically that the tunable agents framework allows easy adaption between cooperative and competitive behaviours in sequential social dilemmas without the need for retraining, allowing a single trained agent model to be adjusted to cater for a wide range of behaviours and opponent strategies.

Tunable Behaviours in Sequential Social Dilemmas using Multi-Objective Reinforcement Learning

Continual learning aims to provide intelligent agents capable of learning multiple tasks sequentially with neural networks. In most settings of the current approaches, the agent starts from randomly initialized parameters and is optimized to master the current task regardless of the usefulness of the learned representation for future tasks. Moreover, each of the future tasks uses all the previously learned knowledge although parts of this knowledge might not be helpful for its learning. These cause interference among tasks, especially when the data of previous tasks is not accessible. In this paper, we propose a new method, named Self-Attention Meta-Learner (SAM), which learns a prior knowledge for continual learning that permits learning a sequence of tasks, while avoiding catastrophic forgetting. SAM incorporates an attention mechanism that learns to select the particular relevant representation for each future task. We empirically show that we can achieve a better performance than several state-of-the-art methods for continual learning by building on top of the selected representation learned by SAM. We also show the role of the meta-attention mechanism in boosting informative features corresponding to the input task and identifying the correct target in the task agnostic inference. Finally, we demonstrate that popular existing continual learning methods gain a performance boost when they adopt SAM as a starting point.

Distributional Monte Carlo Tree Search for Risk-Aware and Multi-Objective Reinforcement Learning

Next from AAMAS 2021

Image Sequence Understanding through Narrative Sensemaking

Similar lecture

No More Hand-Tuning Rewards: Masked Constrained Policy Optimization for Safe Reinforcement Learning

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES