United Kingdom

Policy gradient methods have become one of the most popular classes of algorithms for multi-agent reinforcement learning. A key challenge, however, that is not addressed by many of these methods is multi-agent credit assignment: assessing an agent&#39;s contribution to the overall performance, which is crucial for learning good policies. We propose a novel algorithm called Dr.Reinforce that explicitly tackles this by combining difference rewards with policy gradients to allow for learning decentralized policies when the reward function is known. By differencing the reward function directly, Dr.Reinforce avoids difficulties associated with learning the 
-function as done by Counterfactual Multiagent Policy Gradients (COMA), a state-of-the-art difference rewards method. For applications where the reward function is unknown, we show the effectiveness of a version of Dr.Reinforce that learns a reward network that is used to estimate the difference rewards.

AAMAS 2021

Difference Rewards Policy Gradients

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 

Policy gradient methods have become one of the most popular classes of algorithms for multi-agent reinforcement learning. A key challenge, however, that is not addressed by many of these methods is multi-agent credit assignment: assessing an agent's contribution to the overall performance, which is crucial for learning good policies. We propose a novel algorithm called Dr.Reinforce that explicitly tackles this by combining difference rewards with policy gradients to allow for learning decentralized policies when the reward function is known. By differencing the reward function directly, Dr.Reinforce avoids difficulties associated with learning the 
-function as done by Counterfactual Multiagent Policy Gradients (COMA), a state-of-the-art difference rewards method. For applications where the reward function is unknown, we show the effectiveness of a version of Dr.Reinforce that learns a reward network that is used to estimate the difference rewards.

poster

Automated care systems are becoming more tangible than ever: recent breakthroughs in robotics and machine learning can be used to address the need for automated care created by the increasing aging population. However, such systems require overcoming several technological, ethical, and social challenges. One inspirational manifestation of these challenges can be observed in the training of seeing-eye dogs for visually impaired people. A seeing-eye dog is not just trained to obey its owner, but also to "intelligently disobey": if it is given an unsafe command from its handler, it is taught to disobey it or even insist on a different course of action. This paper proposes the challenge of building a seeing-eye robot, as a thought-provoking use-case that helps identify the challenges to be faced when creating behaviors for robot assistants in general. Through this challenge, this paper delineates the prerequisites that an automated care system will need to have in order to perform intelligent disobedience and to serve as a true agent for its handler.

The Seeing-Eye Robot Grand Challenge: Rethinking Automated Care (Blue Sky Ideas Track)

Human-agent interactions present people guide an object or agent to act as human intentions. This demonstration work develops an online human-agent interaction system, particularly targeting the brain-computer interface (BCI), which uses real-time brain cortex signals: electroencephalogram (EEG) to control the agent in Unity3D game platform. The developed system also provides the online visualisation of EEG signals, including pre-processed temporal data and power spectral in three frequency bands (theta, alpha, and beta). To build this systematic work, we firstly collect wireless EEG signals via the Bluetooth transmission from a commercially available 14-channel brainware headset (Emotiv). EEG signals are then pre-processed and fed into a trained deep learning model to predict the human intentions, which will be sent to Unity3D platform to control an agent’s movements in game playing, such as a karting game scenario. The online testing results show the feasibility of our systematic work that will benefit for human-agent interaction community. The demonstration video can be viewed at the following link: https://youtu.be/9AWKHeatc6I

An Online Human-Agent Interaction System: A Brain-controlled Agent Playing Games in Unity

We propose a generic reward shaping approach for improving the rate of convergence in reinforcement learning (RL), called Self Improvement Based REwards, or SIBRE. The approach is designed for use in conjunction with any existing RL algorithm, and consists of rewarding improvement over the agent's own past performance. We prove that SIBRE converges in expectation under the same conditions as the original RL algorithm. The reshaped rewards help discriminate between policies when the original rewards are weakly discriminated or sparse. Experiments on several well-known benchmark environments with different RL algorithms show that SIBRE converges to the optimal policy faster and more stably. We also perform sensitivity analysis with respect to hyper-parameters, in comparison with baseline RL algorithms.

SIBRE: Self Improvement Based REwards for Adaptive Feedback in 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.

Self-Attention Meta-Learner for Continual Learning

In modular reinforcement learning (MRL), a complex decision making problem is decomposed into multiple simpler subproblems each solved by a separate module. Often, these subproblems have conflicting goals, and incomparable reward scales. A composable decision making architecture requires that even the modules authored separately with possibly misaligned reward scales can be combined coherently. An arbitrator should consider different module's action preferences to learn effective global action selection. We present a novel framework called GRACIAS that assigns fine-grained importance to the different modules based on their relevance in a given state, and enables composable decision making based on modern deep RL methods such as deep deterministic policy gradient (DDPG) and deep Q-learning. We provide insights into the convergence properties of GRACIAS and also show that previous MRL algorithms reduce to special cases of our framework. We experimentally demonstrate on several standard MRL domains that our approach works significantly better than the previous MRL methods, and is highly robust to incomparable reward scales. Our framework extends MRL to complex Atari games such as Qbert, and has a better learning curve than the conventional RL algorithms.

Action Selection for Composable Modular Deep Reinforcement Learning

Symbolic planning models allow decision-making agents to sequence actions in arbitrary ways to achieve a variety of goals in dynamic domains. However, they are typically handcrafted and tend to require precise formulations that are not robust to human error. Reinforcement learning (RL) approaches do not require such models, and instead learn domain dynamics by exploring the environment and collecting rewards. However, RL approaches tend to require millions of episodes of experience and often learn policies that are not easily transferable to other tasks. In this paper, we address one aspect of the open problem of integrating these approaches: how can decision-making agents resolve discrepancies in their symbolic planning models while attempting to accomplish goals? We propose an integrated framework named SPOTTER that uses RL to augment and support ("spot") a planning agent by discovering new operators needed by the agent to accomplish goals that are initially unreachable for the agent. SPOTTER outperforms pure-RL approaches while also discovering transferable symbolic knowledge and does not require supervision, successful plan traces or any a priori  knowledge about the missing planning operator.

SPOTTER: Extending Symbolic Planning Operators through Targeted Reinforcement Learning

Several resource allocation problems involve multiple types of resources, with a different agency being responsible for "locally" allocating the resources of each type, while a central planner wishes to provide a guarantee on the properties of the final allocation given agents' preferences. We study the relationship between properties of the local mechanisms, each responsible for assigning all of the resources of a designated type, and the properties of a sequential mechanism which is composed of these local mechanisms, one for each type, applied sequentially, under lexicographic preferences, a well studied model of preferences over multiple types of resources in artificial intelligence and economics. We show that when preferences are O-legal, meaning that agents share a common importance order on the types, sequential mechanisms satisfy the desirable properties of anonymity, neutrality, non-bossiness, or Pareto-optimality if and only if every local mechanism also satisfies the same property, and they are applied sequentially according to the order O. Our main results are that under O-legal lexicographic preferences, every mechanism satisfying strategyproofness and a combination of these properties must be a sequential composition of local mechanisms that are also strategyproof, and satisfy the same combinations of properties.

Sequential Mechanisms for Multi-type Resource Allocation

To help facilitate play and learning, game-based educational activities often feature a computational agent as a co-player. Personalizing this agent's behavior to the student player is an active area of research, and prior work has demonstrated the benefits of personalized educational interaction across a variety of domains. A critical research challenge for personalized educational agents is real-time student modeling. Most student models are designed for and trained on only a single task, which limits the variety, flexibility, and efficiency of student player model learning.    In this paper we present a research project applying transfer learning methods to student player models over different educational tasks, studying the effects of an algorithmic "multi-task personalization" approach on the accuracy and data efficiency of student model learning. We describe a unified robotic game system for studying multi-task personalization over two different educational games, each emphasizing early language and literacy skills such as rhyming and spelling. We present a flexible Gaussian Process-based approach for rapidly learning student models from interactive play in each game, and a method for transferring each game's learned student model to the other via a novel instance-weighting protocol based on task similarity. We present results from a simulation-based investigation of the impact of multi-task personalization, establishing the core viability and benefits of transferrable student models and outlining new questions for future in-person research.

Towards Transferrable Personalized Student Models in Educational Games

People often encounter high-stress situations. Modeling and being able to predict people’s behavior in such situations, how they cope, is a critical research topic. To that end, we propose a computational model of coping that casts Lazarus' theory of coping into a Partial Observable Markov Decision Process (POMDP) framework. This includes an appraisal process that models the factors that lead to stress by assessing a person's relation to the environment and a coping process that models people’s behavior in the face of such stress. This coping process includes problem-focused coping, whereby people seek to alter the external environment, and emotion-focused coping, whereby people alter their internal beliefs, goals, and intentions in the face of stress. We evaluate the model’s assumptions and predictions in the context of a high-stress situation that is increasingly common, the extreme conditions of a hurricane. We collected human survey data from the last several years of major U.S. hurricanes to evaluate the features in the models used for appraisal calculation. Additionally, we conducted a controlled human-subject experiment simulating a hurricane experience to investigate the prediction of the model on how people change their beliefs and goals to cope with the situation. The results show that, as predicted by the model, the proposed model features are significantly associated with the evacuation decisions and post-decision people also change their beliefs and goals in the directions that align with their prior decisions. Lastly, we conduct a simulation study showing that the proposed model is qualitatively closer to the experiment data than the baseline models that do not incorporate coping effects.

A Computational Model of Coping for Simulating Human Behavior in High-Stress Situations

Agents operating in the open world often produce negative side effects (NSE), which are difficult to identify at design time. We examine how a human can assist an agent, beyond providing feedback, and exploit their broader scope of knowledge to mitigate the impacts of NSE. We formulate this problem as a human-agent team with decoupled objectives. The agent optimizes its assigned task, during which its actions may produce NSE. The human shapes the environment through minor reconfiguration actions so as to mitigate the impacts of agent's side effects, without significantly degrading agent performance. We present an algorithm to solve this problem. Empirical evaluation shows that the proposed framework can successfully mitigate NSE, without affecting the agent's ability to complete its assigned task.

Difference Rewards Policy Gradients

Next from AAMAS 2021

The Seeing-Eye Robot Grand Challenge: Rethinking Automated Care (Blue Sky Ideas Track)

Similar lecture

Learning Node-Selection Strategies in Bounded-Suboptimal Conflict-Based Search for Multi-Agent Path Finding