New Zealand

This paper tackles the problem of learning generalizable congestion-mitigation strategies in simple representations of traffic. In particular, we look to mixed-autonomy ring roads as depictions of instabilities common to many generic settings, and ask the question: What features are needed to ensure that policies here can be adapted to typical multi-lane highways? To answer this, we study the implications of the scale of the source task and the modeling of pseudo-lane change events within it on the transferability of policies learned to complex networks. Our findings suggest that negating the effects of boundary conditions and introducing lane changes that approximately match trends in more complex systems can significantly improve the generalizability of learned behaviors.

AAMAS 2022

Learning Generalizable Multi-Lane Mixed Autonomy Control Strategies in Single-Lane Settings

traffic control

social simulation

reinforcement learning

AAMAS (International Conference on Autonomous Agents and Multiagent Systems) is the largest and most influential conference in the area of agents and multiagent systems. The aim of the conference is to bring together researchers and practitioners in all areas of agent technology and to provide a single, high-profile, internationally renowned forum for research in the theory and practice of autonomous agents and multiagent systems. AAMAS is the flagship conference of the non-profit International Foundation for Autonomous Agents and Multiagent Systems (IFAAMAS).

The AAMAS conference series was initiated in 2002 in Bologna, Italy as a joint event comprising the 6th International Conference on Autonomous Agents (AA), the 5th International Conference on Multiagent Systems (ICMAS), and the 9th International Workshop on Agent Theories, Architectures, and Languages (ATAL).

Subsequent AAMAS conferences have been held in Melbourne, Australia (July 2003), New York City, NY, USA (July 2004), Utrecht, The Netherlands (July 2005), Hakodate, Japan (May 2006), Honolulu, Hawaii, USA (May 2007), Estoril, Portugal (May 2008), Budapest, Hungary (May 2009), Toronto, Canada (May 2010), Taipei, Taiwan (May 2011), Valencia, Spain (June 2012), Minnesota, USA (May 2013), Paris, France (May 2014), Istanbul, Turkey (May 2015), Singapore (May 2016), São Paulo (2017) and Stockholm, Sweden (2018), Montreal (May 2019), Auckland (May 2020, Virtual), London (May 2021, Virtual).
<br>
<br>



A ticket is required to attend this event, please register using the link below:

https://aamas2022-conference.auckland.ac.nz/attending/registration/

Registration Is Required

AAMAS (International Conference on Autonomous Agents and Multiagent Systems) is the largest and most influential conference in the area of agents and multiagent systems. The aim of the conference is to bring together researchers and practitioners in all areas of agent technology and to provide a single, high-profile, internationally renowned forum for research in the theory and practice of autonomous agents and multiagent systems. AAMAS is the flagship conference of the non-profit International Foundation for Autonomous Agents and Multiagent Systems (IFAAMAS).

poster


The Belief Desire Intention model is a widely used architecture for
developing rational agents. Because of its expressiveness, the task
of programming a BDI agent can be challenging, especially when
applied to safety-critical scenarios. In such scenarios, it is important
to provide a safeguard for the critical behaviour of the agent. In this
poster, we summarise how to extend the agent’s reasoning cycle
in the BDI model with safety shields. A safety shield works as a
sandbox for the agents’ plans that is enforced at runtime so that the
agent behaves according to a safety formal specification. A runtime
monitor is automatically synthesised from a shield to detect any
failure that is within the scope of the shielded plan.

Safety Shields, an Automated Failure Handling Mechanism for BDI Agents

We introduce the Forecasting Argumentation Framework (FAF), a novel argumentation framework for forecasting informed by recent judgmental forecasting research. FAFs comprise update frameworks which empower (human or artificial) agents to argue over time with and about probability of scenarios, whilst flagging perceived irrationality in their behaviour with a view to improving their forecasting accuracy. FAFs include three argument types with future forecasts and aggregate the strength of these arguments to inform estimates of the likelihood of scenarios. We provide an example of an implementation of FAFs for supporting forecasting agents.

Argumentative Forecasting


In sequential multi-objective decision making (MODeM) settings, when the utility of a user is derived from a single execution of a policy, policies for the expected scalarised returns (ESR) criterion should be computed. In multi-objective settings, a user's preferences over objectives, or utility function, may be unknown at the time of planning. When the utility function of a user is unknown, multi-policy methods are deployed to compute a set of optimal policies. However, the state-of-the-art sequential MODeM multi-policy algorithms compute a set of optimal policies for the scalarised expected returns (SER) criterion. Algorithms that compute a set of optimal policies for the SER criterion utilise expected value vectors which cannot be used when optimising for the ESR criterion. We propose multi-objective distributional value iteration (MODVI) that replaces value vectors with distributions over the returns and computes a set of optimal policies for the ESR criterion.

Decision-Theoretic Planning for the Expected Scalarised Returns

In adversarial patrolling games, a mobile Defender strives to discover intrusions at vulnerable targets initiated by an Attacker.
The Attacker's utility is traditionally defined as the probability of completing an attack, possibly weighted by target costs.
However, in many real-world scenarios, the actual damage caused by the Attacker depends on the time elapsed since the attack's initiation to its detection.
We introduce a formal model for such scenarios, and we show that the Defender always has an optimal strategy achieving maximal protection.
We also prove that finite-memory Defender's strategies are sufficient for achieving protection arbitrarily close to the optimum.
Then, we design an efficient strategy synthesis algorithm based on differentiable programming and gradient descent.
We evaluate the efficiency of our method experimentally.

Minimizing Expected Intrusion Detection Time in Adversarial Patrolling

We consider the computational problem of partitioning items into bundles among alternatives to maximize social welfare. Unfortunately, many important classes of this problem are computationally hard, including well-known instances in the multi-agent systems literature. We analyze novel concise representations and restrictions that admit polynomial-time algorithms for many such combinatorial assignment problems, and prove several complexity results for them. We provide efficient approximation algorithms and non-trivial exponential-time algorithms for the hard cases.

Concise Representations and Complexity of Combinatorial Assignment Problems

We consider the popular Partially Observable Monte-Carlo Planning (POMCP) algorithm and propose a methodology, called Active XPOMCP, for generating compact logical rules that represent properties of the control policy.
These rules are then used as shields to prevent POMCP from selecting unexpected actions, with useful implications on the security and trustworthiness of the algorithm.
Contrary to state-of-the-art methods, Active XPOMCP does not require a previously generated set of belief-action pairs to generate the logical rule, but it actively generates this data in an information-efficient way by querying the algorithm.
Active XPOMCP reduces the number of beliefs needed to generate accurate rules with respect to state-of-the-art methods, and it allows to produce more accurate shields when few belief-action samples are available.

Active Generation of Logical Rules for POMCP Shielding


We study the problem of fair allocation of a set of indivisible items among agents with additive valuations, under cardinality constraints. In this setting, the items are partitioned into categories, each with its own limit on the number of items it may contribute to any bundle. We consider the fairness measure known as the maximin share (MMS) guarantee and propose a novel polynomial-time algorithm for finding 1/2-approximate MMS allocations–an improvement from the previously best available guarantee of 11/30.

Guaranteeing Half-Maximin Shares Under Cardinality Constraints


Multi-agent task allocation methods seek to distribute a set of tasks fairly amongst a set of agents. In real-world settings, such as fruit farms, human labourers undertake harvesting tasks, organised each day by farm manager(s) who assign workers to the fields that are ready to be harvested. The work presented here considers three challenges identified in the adaptation of a multi-agent task allocation methodology applied to the problem of distributing workers to fields. First, the methodology must be fast to compute so that it can be applied on a daily basis. Second, the incremental acquisition of harvesting data used to make decisions about worker-task assignments means that a data-backed approach must be derived from incomplete information as the growing season unfolds. Third, the allocation must take “fairness” into account and consider worker motivation. Solutions to these challenges are demonstrated, showing statistically significant results based on the operations at a soft fruit farm during their 2020 and 2021 harvesting seasons.

Multi-agent Task Allocation for Fruit Picker Team Formation


Training reinforcement learning agents in real-world environments is costly, particularly for safety-critical applications. Human input can enable an agent to learn a good policy while avoiding unsafe actions, but at the cost of bothering the human with repeated queries. We present a model for safe learning in safety-critical environments from human input that minimises bother cost. Our model, JPAL-HA, proposes an efficient mechanism to harness human preferences and justifications to significantly improve safety during the learning process without increasing the number of interactions with a user. We show this with both simulation and human experiments.

How to train your agent: Active Learning from Human Preferences and Justifications in Safety-critical Environments


Affective reaction is a mechanism that is triggered by a perceived emotion and results in the observer's emotion, subsequently steering the behavior of the observer. It is crucial to understand this mechanism as it constitutes a mediator between the emotional expressions of one agent and the (behavioral) responses of another thus facilitating social interaction among agents. This paper reports on a formal specification of affective reaction based on its psychological analysis using values and value systems as agents' appraisal standards.

Learning Generalizable Multi-Lane Mixed Autonomy Control Strategies in Single-Lane Settings

Downloads

Next from AAMAS 2022

Safety Shields, an Automated Failure Handling Mechanism for BDI Agents

Similar lecture

Cooperation and learning dynamics under risk diversity and financial incentives

Learning Generalizable Multi-Lane Mixed Autonomy Control Strategies in Single-Lane Settings

.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 2022

Safety Shields, an Automated Failure Handling Mechanism for BDI Agents

Similar lecture

Cooperation and learning dynamics under risk diversity and financial incentives

Downloads