United Kingdom

We present the first explicit-state method for analysing and ensuring the safety of DRL agents for Atari games. Our method only requires access to the emulator. We give a suite of 42 properties that characterise &quot;safe behaviour&quot; for 31 games. We evaluate the safety of the best available DRL agents which, as our experiments show, violate most of our properties. We propose a countermeasure that implements shielding using bounded explicit-state exploration. Our method improved their overall safety, producing the safest DRL agents for Atari games currently available.

AAMAS 2021

Shielding Atari Games with Bounded Prescience

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 

We present the first explicit-state method for analysing and ensuring the safety of DRL agents for Atari games. Our method only requires access to the emulator. We give a suite of 42 properties that characterise "safe behaviour" for 31 games. We evaluate the safety of the best available DRL agents which, as our experiments show, violate most of our properties. We propose a countermeasure that implements shielding using bounded explicit-state exploration. Our method improved their overall safety, producing the safest DRL agents for Atari games currently available.

poster

We model interpretive blindness, a type of epistemic bias that poses a problem for learning from testimony, in which one acquires information from text or conversation but lacks direct access to ground truth. Interpretive blindness arises when a co-dependence between background beliefs and interpretation leads to a dynamic process of bias hardening that impedes or precludes learning. We argue that when bodies of data are argumentatively complete, even constraints from hierarchical Bayesian learning designed to promote good epistemic practices will fail to stop interpretive blindness.

Interpretive Blindness and the Impossibility of Learning from Testimony

In this work we introduce an interactive machine teaching approach that teaches a classification task to the learner. Our adaptive approach - Feature Importance Teaching (FIT) - does not assume perfect knowledge about the learner, as most machine teaching approaches do. It chooses, online, which sample to show next, as it updates the learner's model based on feedback from the student on the weights attributed to the features. We present simulated results where the student has a different prior knowledge from the one assumed by the teacher. The results have shown that our teaching approach can mitigate this mismatch and lead to a significantly faster learning curve than the ones obtained in conditions where the teacher randomly selects the samples or does not consider this feedback from the student.

Teaching Unknown Learners to Classify via Feature Importance

Recently, there is growing attention on applying deep reinforcement learning (DRL) to solve the 3D bin packing problem (3D BPP), given its favorable generalization and independence of ground-truth label. However, due to the relatively less informative yet computationally heavy encoder, and considerably large action space inherent to the 3D BPP, existing methods are only able to handle up to 50 boxes. In this paper, we propose to alleviate this issue via an end-to-end multimodal DRL agent, which sequentially addresses three sub-tasks of sequence, orientation and position, respectively. The resulting architecture enables the agent to solve large-scale instances of 100 boxes or more. Experiments show that the agent could learn highly efficient policies that deliver superior performance against all the baselines on instances of various scales.

Solving 3D Bin Packing Problem via Multimodal Deep Reinforcement Learning

This paper is concerned with designing purpose in autonomous robots for open-ended learning settings. Unconstrained human robot interaction situations and robotic systems that must operate in dynamic multi-robot scenarios are paradigmatic examples of open-endedness. An approach to the appropriate design and engineering of motivational structures to endow robots with a particular purpose is proposed and tested. This approach focuses on the drive structure and how it can be made to autonomously adapt to changing circumstances. Specifically, a simple evolutionary strategy for the autonomous regulation of multiple drives in order to optimize long-term operation is defined. The experimental results have been obtained on a Baxter robot facing changing situations in real setups.

An Autonomous Drive Balancing Strategy for the Design of Purpose in Open-ended Learning Robots

Most existing incentive allocation approaches rely on sufficient information about users' attributes, such as their preferences, followers in the social network, and activities, to customize effective incentives. However, this may lead to failure when such knowledge is unavailable. In this light, we propose an end-to-end reinforcement learning-based framework, named Geometric Actor-Critic (GAC), to discover effective incentive allocation policies towards users in a social network. More specifically, given a limited budget, the proposed approach can extract information from a high-level network representation for learning effective incentive allocation policies. The proposed GAC only requires the topology of the social network and does not rely on any prior information about users' attributes. We use three real-world social network datasets to evaluate the performance of the proposed GAC. The experimental results demonstrate the effectiveness of the proposed approach.

Learning Policies for Effective Incentive Allocation in Unknown Social Networks

Several frameworks for decentralized reasoning assume a junction tree agent organization (JT-org). JT-org construction involves 3 related tasks on existence recognition, construction, and environment re-decomposition, where re-decomposition incurs loss of JT-org linked privacy, including privacy on agent, topology, private and shared variables. We propose a novel algorithm DAER that accomplishes all 3 tasks distributively. For Tasks 1 and 2, DAER incurs no loss of JT-org linked privacy.For Task 3, it incurs significantly less privacy loss than existing JT-org construction methods. Its performance is formally analyzed and empirically evaluated.

Constructing Junction Tree Agent Organization with Privacy (JAAMAS Track)

Creating incentives for cooperation is a challenge in natural and artificial systems.   One potential answer is reputation, whereby agents trade the immediate cost of cooperation for the future benefits of having a good reputation.   Game theoretical models have shown that specific social norms can make cooperation stable, but how agents can independently learn to establish effective reputation mechanisms on their own is less understood. We use a simple model of reinforcement learning to show that reputation mechanisms generate two coordination problems: agents need to learn how to coordinate on the meaning of existing reputations and collectively agree on a social norm to assign reputations to others based on their behavior. These coordination problems exhibit multiple equilibria, some of which effectively establish cooperation. When we train agents with a standard Q-learning algorithm in an environment with the presence of reputation mechanisms, convergence to undesirable equilibria is widespread. We propose two mechanisms to alleviate this: (i) seeding a proportion of the system with fixed agents that steer others towards good equilibria; and (ii), intrinsic rewards based on the idea of introspection, i.e., augmenting agents' rewards by an amount proportionate to the performance of their own strategy against themselves. A combination of these simple mechanisms is successful in stabilizing cooperation, even in a fully decentralized version of the problem where agents learn to use and assign reputations simultaneously. We show how our results relate to the literature in Evolutionary Game Theory, and discuss implications for artificial, human and hybrid systems, where reputations can be used as a way to establish trust and cooperation.

Cooperation and Reputation Dynamics with Reinforcement Learning

Can artificial agents benefit from human conventions? Human societies manage to successfully self-organize and resolve the tragedy of the commons in common-pool resources, in spite of the bleak prediction of non-cooperative game theory. On top of that, real-world problems are inherently large-scale and of low observability. One key concept that facilitates human coordination in such settings is the use of conventions. Inspired by human behavior, we investigate the learning dynamics and emergence of temporal conventions, focusing on common-pool resources. Extra emphasis was given in designing a realistic evaluation setting: (a) environment dynamics are modeled on real-world fisheries, (b) we assume decentralized learning, where agents can observe only their own history, and (c) we run large-scale simulations (up to 64 agents).

Improved Cooperation by Exploiting a Common Signal

Reasoning about observed effects and their causes is important in many applications. For instance, understanding why a plan failed can aid the task of replanning by allowing the agent to tailor a better plan. But under incomplete information, an agent may be unable to determine which actions/events caused an effect. To overcome this, the agent may be able to perform some sensing actions that allow him to figure out what caused the effect. This becomes even more important in multiagent contexts, where an agent may want to identify which agents caused some effect, or possibly prevent other agents from determining who caused something. The effects involved may even be epistemic effects, such as an agent coming to know the PIN of a bank card, and the causes may be sensing actions. Reasoning about such causes is a key part of "theory of mind" and understanding other agents' behaviour. While there has been much work on causality from an objective standpoint, causality from the point of view of individual agents has received much less attention. In this paper, we develop a formalization of knowledge about actual causes in the situation calculus, and how it is affected by actions including sensing. We show that the proposed framework has some intuitive properties and study the conditions under which an agent can be expected to come to know the causes of an effect.

Knowing Why - On the Dynamics of Knowledge about Actual Causes in the Situation Calculus

High-performing teams learn effective communication strategies to judiciously share information and reduce the cost of communication overhead. Within multi-agent reinforcement learning, synthesizing effective policies requires reasoning about when to communicate, whom to communicate with, and how to process messages. We propose a novel multi-agent reinforcement learning algorithm, Multi-Agent Graph-attentIon Communication (MAGIC), with a graph-attention communication protocol in which we learn 1) a Scheduler to help with the problems of when to communicate and whom to address messages to, and 2) a Message Processor using Graph Attention Networks (GATs) with dynamic graphs to deal with communication signals. The Scheduler consists of a graph attention encoder and a differentiable attention mechanism, which outputs dynamic, differentiable graphs to the Message Processor, which enables the Scheduler and Message Processor to be trained end-to-end. We evaluate our approach on a variety of cooperative tasks, including Google Research Football. Our method outperforms baselines across all domains, achieving ~10.5% increase in reward in the most challenging domain. We also show MAGIC communicates $27.4%$ more efficiently on average than baselines, is robust to stochasticity, and scales to larger state-action spaces. Finally, we demonstrate MAGIC on a physical, multi-robot testbed.

Shielding Atari Games with Bounded Prescience

Next from AAMAS 2021

Interpretive Blindness and the Impossibility of Learning from Testimony

Similar lecture

Eliciting Fairness in Multiplayer Bargaining through Network-Based Role Assignment

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES