PURE-MAS

Planning under uncertainty for real-world multiagent systems

 Partecipanti

Mappa

 Word cloud

Esplora la "nuvola delle parole (Word Cloud) per avere un'idea di massima del progetto.

 Obiettivo del progetto (Objective)

'A major goal of Artificial Intelligence is designing agents: systems that perceive their environment and execute actions. In particular, a fundamental question is how to build intelligent agents. When uncertainty and many agents are involved, this question is particularly challenging, and has not yet been answered in a satisfactory way.

The need for scalable and flexible multiagent planning is particularly pressing given that intelligent distributed systems are becoming ubiquitous in society. For instance, autonomous guided vehicles transport cargo and people, inter-vehicle communication allow cars to form vehicular networks, smart grid infrastructure allows consumers to produce and sell electricity, and surveillance cameras provide urban security and safety. In these settings, the controller of the vehicle, the consumers, or the controller of the cameras all need to act in the face of uncertainty.

For an agent in isolation, planning under uncertainty has been studied using decision-theoretic models like Partially Observable Markov Decision Processes (POMDPs). Such single-agent, centralized methods clearly do not suffice for large-scale multiagent systems. I focus on multiagent techniques, and I propose to advance the state of the art as follows.

First, instead of the rigid history-based plans currently in use, I will develop a flexible plan representation with methods for determining the impact on other agents of updating one agent's plan. Second, I will consider scenarios with self-interested agents, relevant in domains such as smart grids or cars driving on a highway. I will be able to scale up non-cooperative techniques by exploiting local interactions between agents. These advances will be empirically tested in intelligent transportation systems and smart grids.

Moving to TU Delft will allow me to learn from their expertise on re-planning, and on planning with self-interested agents, as well as provide unique opportunities for the empirical evaluation.'

Descrizione progetto (Article)

From autonomous vehicles to vehicular ad hoc networks or smart grids, intelligent distributed systems can be found everywhere. For an isolated agent, planning under uncertainty has already been thoroughly studied. However, centralised methods for single agents clearly do not suffice for large-scale multi-agent systems.

Funded by the EU, the 'Planning under uncertainty for real-world multi-agent systems' (PURE-MAS) project addressed the issue of the multi-agent decision process. Research was based on a variety of multi-agent models for cooperative agents. These included decentralised partially observable Markov decision process (dec-POMDP) and other related models known as multi-agent sequential decision making.

Scientists improved scalability both in terms of planning horizon and number of agents. By proposing one of the fastest optimal dec-POMDP planners, scientists were able to compute previously unattainable planning horizons. Furthermore, through exploiting their structure, important advances were made for scaling up planning to unprecedented team sizes.

Besides scalability, PURE-MAS explored two types of event-based representations for capturing relevant changes in the environment, and as such, providing a higher-level abstraction. This paves the way for online planning agents that can adapt to the dynamic environment by continuously modifying plans during plan execution. Several algorithms were also proposed for optimising communication between agents.

By combining game theory and planning, the project team achieved coordinated planning amongst single agents. In particular, techniques drawn from dynamic mechanism design helped align the incentives of multiple contractors. Based on these techniques, human players can learn that cooperation amongst competitors can be beneficial and road authorities can draw up better maintenance contracts.

Another project achievement was to demonstrate a small team of robots that could successfully track a mobile target. PURE-MAS also optimised the multi-agent system to provide intelligence to a distributed smart grid.

PURE-MAS contributed significantly to achieving one of the major goals of artificial intelligence, namely how to build many intelligent agents that perceive their environment and execute appropriate actions. All project advances can find applications in transportation systems, smart grids and surveillance cameras.