Projects
Main.Projects History
Hide minor edits - Show changes to markup
description to be added
Autonomous agents transcend their individual capabilities by cooperating towards achieving shared goals. In complex multiagent environments, agents may have different viewpoints due to locality of sensing, different fundamental assumptions about the domain, or simply because different agents may have conducted different inferences and therefore their beliefs may not be aligned. In this project we aim to develp an argumentation-based dialogue framework that allows agents to discuss candidate plans, align their viewpoints and reach agreements.
description to be added
In multiagent systems it is often desirable to achieve coordination through the use of restrictions on agent behaviour. These restrictions, or social norms, conditionally prohibit agent action so as to achieve a social objective. In this project, we are interested in algorithmic processes that can be used to synthesise these restrictions. Our approach adopts a planning-based domain formalism to synthesise norms, and utilises planning itself to investigate the effects the synthesised norms have on the goals of the agents.
In multiagent systems it is often desirable to achieve coordination through the use of restrictions on agent behaviour. These restrictions, or social norms, conditionally prohibit agent action so as to achieve a social objective. In this project, we are interested in algorithmic processes that can be used to synthesise these restrictions. Our approach adopts a planning-based domain formalism to synthesise norms, and utilises planning itself to investigate what effects the synthesised norms have on the goals of the agents.
description to be added
In multiagent systems it is often desirable to achieve coordination through the use of restrictions on agent behaviour. These restrictions, or social norms, conditionally prohibit agent action so as to achieve a social objective. In this project, we are interested in algorithmic processes that can be used to synthesise these restrictions. Our approach adopts a planning-based domain formalism to synthesise norms, and utilises planning itself to investigate the effects the synthesised norms have on the goals of the agents.
description to be added
This project looks at multiagent planning domains with self-interested agents. In particular, a parcel delivery gridworld domain is implemented. This domain is highly coupled in terms of the possible interactions between agents and contains interesting scenarios for cooperation. Game Theoretic concepts are applied to develop algorithms for finding stable solutions (joint plans).
description to be added
Modelling learning opponents in games (since 2007)
Francesco Figari, Michael Rovatsos, Ed Hopkins
Vagueness (since 2009)
Ewan Klein, Michael Rovatsos
In this project, we are interested in the practical benefits of using vague terms in multiagent communication. We are currently developing a simulation testbed within which we can experiment with different vague semantics in multiagent communication in a simple application domain, and hope that this will shed more light on the impact of vagueness on the adaptability, robustness, and scalability of evolving semantics.
Empirical semantics (2003-2006)
Matthias Nickles, Michael Rovatsos, Gerhard Weiss, Wilfried Brauer
The aim of this project was to give an alternative characterisation for the semantics of agent communication that was based on expectations based on experience rather than on mental models of other agents or on publicly visible commitments. We developed a probabilistic framework that included both system-level representations of the evolving meaning of expectation-based semantics, and agent-level simplified representations of meaning that can be used for making actual communication decisions from the perspective of an individual agent. A summary of this work can be found in (Nickles et al 2007).
Dynamic semantics (2006-2007)
Michael Rovatsos
Building on the empirical semantics framework, we developed a logic-based mechanism for tracking concrete plan-based semantics of speech acts in a commitment-based agent communication setting. The contribution of this work is that it allows for an explicit representation of different "states" of the meaning of certain utterances where transitions between different states are contingent on agent behaviour. The model is described in (Rovatsos 2007).
In this project, we applied hierarchical reinforcement learning techniques to symbolic agent communication languages to enable agents to learn time-extended policies in a strategic communication environment. (Fischer, Rovatsos and Weiss 2005) and (Fischer and Rovatsos 2005) are the main references for this work. Later, we also applied this framework to argumentation-based negotiation (Rovatsos et al 2005); to our knowledge, this was the first work to attempt online learning of argumentation strategies in multiagent dialogues.
In this project, we applied hierarchical reinforcement learning techniques to symbolic agent communication languages to enable agents to learn time-extended policies in a strategic communication environment. (Fischer et al 2005) and (Fischer and Rovatsos 2005) are the main references for this work. Later, we also applied this framework to argumentation-based negotiation (Rovatsos et al 2005); to our knowledge, this was the first work to attempt online learning of argumentation strategies in multiagent dialogues.
Interaction frames and framing architecture (2000-2004)
Michael Rovatsos, Kai Paetow, Felix Fischer, Gerhard Weiss
The aim of this project was to design an architecture for managing strategic interactions between agents based on concepts of "frames" and "framing" as suggested by the sociologist Erving Goffman. The resulting framework provided a holistic method for reasoning about agent-to-agent communicative interactions and for learning how to abstract from individual interaction experiences and to optimise interaction choices in the long-term. An overview of the framework is given in (Rovatsos et al 2002)
Autonomy modelling (2001-2004)
Gerhard Weiss, Matthias Nickles, Michael Rovatsos, Christian Meindl
This work developed languages and theoretical models for describing autonomy relationships between interacting agents and how these change with communication. We developed both a conceptual framework that allows for viewing autonomy as a communication-mediated variable property of social systems, and concrete formal languages that allow for the specification of deontic relationships and how these change over time. See (Rovatsos and Weiss 2005) and (Weiss et al 2005).
Expectation-oriented analysis and design (2000-2004)
Matthias Nickles, Michael Rovatsos, Gerhard Weiss, Wilfried Brauer
The goal of this project was to develop a software engineering methodology based on modelling the evolution of expectations among interacting agents in agent-based systems. In contrast to the more theoretical work on empirical semantics, here we suggested a concrete set of guidelines for representing, tracking and improving systems of changing expectations as a suitable method for engineering open systems in which it is hard to predict agent behaviour at design time. The results of this work are discussed in (Nickles et al 2005).
This project was concerned with learning how to dynamically "stereotype" opponents in games with many opponents into different classes based on their strategies. We developed the AdHoc heuristic that was later on used by other researchers and applied to such domains as robotic soccer. Relevant publications are (Rovatsos and Wolf 2002) and (Rovatsos, Weiss, and Wolf 2003). We also developed methods for improving player satisfaction in video games based on classifying human opponents according to their level of skill at the game and adapting to this level (Collins and Rovatsos, 2006).
This project was concerned with learning how to dynamically "stereotype" opponents in games with many opponents into different classes based on their strategies. We developed the AdHoc heuristic that was later on used by other researchers and applied to such domains as robotic soccer. Relevant publications are (Rovatsos and Wolf 2002) and (Rovatsos et al 2003). We also developed methods for improving player satisfaction in video games based on classifying human opponents according to their level of skill at the game and adapting to this level (Collins and Rovatsos, 2006).
Early work on this topic was concerned with integrating witness information into models of multiagent trust, and was the first model to suggest the use of such information. The model is described in one of the most-cited early papers on trust and reputation (Schillo, Rovatsos and Funk 2000)
Early work on this topic was concerned with integrating witness information into models of multiagent trust, and was the first model to suggest the use of such information. The model is described in one of the most-cited early papers on trust and reputation (Schillo et al 2000)
We work on various topics in agents and multiagent systems, mostly related to reasoning about interaction, whether this means modelling other agents, trying to resolve conflicts among interacting agents or learning what governs other's behaviour.
The following visualisation shows them along two critical dimensions, cooperative vs. adversarial and single-agent vs. multi-agent. As much as possible, we've attempted to represent similarity between different themes by proximity (although it is not always possible in 2D). Click on this map to find out more about the individual projects (feature not available yet).
We work on various topics in agents and multiagent systems, mostly related to reasoning about interaction, whether this means modelling other agents, trying to resolve conflicts among interacting agents or learning what governs other's behaviour.
The following visualisation shows work done by members of the group along two critical dimensions, cooperative vs. adversarial and single-agent vs. multi-agent. We've attempted to capture similarities between different themes by proximity to the extent that this is possible in a two-dimensional map. Below the map you will find descriptions of individual research projects (feature not available yet).
Planning games (since 2009)
Matt Crosby, Michael Rovatsos
description to be added
Automated norm synthesis (since 2007)
George Christelis, Michael Rovatsos, Ron Petrick
description to be added
Argumentation in planning (since 2007)
Alexandros-Sotiris Belesiotis, Iyad Rahwan, Guillermo Simari, Michael Rovatsos
description to be added
Expectation-strategy-behaviour architecture (since 2006)
Iain Wallace, Michael Rovatsos
description to be added
Matthew Whitaker, Alexandros Belesiotis, Michael Rovatsos, Jad Hamza
Matthew Whitaker, Alexandros-Sotiris Belesiotis, Jad Hamza, Michael Rovatsos
Michael Rovatsos, Jan Tozicka, Xavier Rafael Palou, Michael Pechoucek
Jan Tozicka, Xavier Rafael Palou, Michael Rovatsos, Michael Pechoucek
Michael Rovatsos, Felix Fischer, Iyad Rahwan, Liz Sonenberg, Gerhard Weiss
In this project, we applied hierarchical reinforcement learning techniques to symbolic agent communication languages to enable agents to learn time-extended policies in a strategic communication environment. (Fischer, Rovatsos and Weiss 2005) and (Fischer and Rovatsos 2005) are the main references for this work. Later, we also applied this framework to argumentation-based negotiation (Rovatsos et al 2005); to our knowledge, this was the first work that attempted to enable online agent-side learning of argumentation strategies.
Felix Fischer, Michael Rovatsos, Iyad Rahwan, Liz Sonenberg, Gerhard Weiss
In this project, we applied hierarchical reinforcement learning techniques to symbolic agent communication languages to enable agents to learn time-extended policies in a strategic communication environment. (Fischer, Rovatsos and Weiss 2005) and (Fischer and Rovatsos 2005) are the main references for this work. Later, we also applied this framework to argumentation-based negotiation (Rovatsos et al 2005); to our knowledge, this was the first work to attempt online learning of argumentation strategies in multiagent dialogues.
Michael Rovatsos, Marco Wolf, Gerhard Weiss, Brian Collins
Marco Wolf, Brian Collins, Michael Rovatsos, Gerhard Weiss
Trust and reputation (since 1999)
Michael Schillo, Michael Rovatsos, Petra Funk, Sebastian Mhatre, Mariola Mier
Early work on this topic was concerned with integrating witness information into models of multiagent trust, and was the first model to suggest the use of such information. The model is described in one of the most-cited early papers on trust and reputation (Schillo, Rovatsos and Funk 2000) Later work looked at estimating the strategic value of lying in repeated games, and currently we are interested in hierarchical notions of trust (e.g. intra- vs inter-group trust).
Communication in reinforcement learning (since 2006, ongoing)
Communication in reinforcement learning (since 2005)
Communication in reinforcement learning (since 2006, ongoing)
Matthew Whitaker, Alexandros Belesiotis, Michael Rovatsos, Jad Hamza
In this project we are interested in two aspects of how communication can be used in multiagent reinforcement learning: (1) as potentially unreliable advice from an information source that makes suggestions to the agents as to how they could better cooperate; and (2) as signals exchanged by the learners while learning, either for the purposes of communicating part of their own state to overcome partial observability, or to communicate their intended actions.
Collaborative learning (2005-2008)
Michael Rovatsos, Jan Tozicka, Xavier Rafael Palou, Michael Pechoucek
This work resulted in an abstract framework for collaborative agent-based machine learning that allows for the integration of heterogeneous agent learning algorithms and enables knowledge sharing among learners in the form of data and hypothesis sharing (Tozicka et 2008). The advantage of our architecture is that it provides a broader range of potential interactions between collaborating learners than previous approaches. The framework has been successfully applied to different real-world learning problems such as brain tumor classification and vessel tracking.
Learning dialogue strategies (2000-2005)
Michael Rovatsos, Felix Fischer, Iyad Rahwan, Liz Sonenberg, Gerhard Weiss
In this project, we applied hierarchical reinforcement learning techniques to symbolic agent communication languages to enable agents to learn time-extended policies in a strategic communication environment. (Fischer, Rovatsos and Weiss 2005) and (Fischer and Rovatsos 2005) are the main references for this work. Later, we also applied this framework to argumentation-based negotiation (Rovatsos et al 2005); to our knowledge, this was the first work that attempted to enable online agent-side learning of argumentation strategies.
Learning dialogue strategies (2000-2005)
Michael Rovatsos, Felix Fischer, Iyad Rahwan, Liz Sonenberg, Gerhard Weiss
In this project, we applied hierarchical reinforcement learning techniques to symbolic agent communication languages to enable agents to learn time-extended policies in a strategic communication environment. (Fischer, Rovatsos and Weiss 2005) and (Fischer and Rovatsos 2005) are the main references for this work. Later, we also applied this framework to argumentation-based negotiation (Rovatsos et al 2005); to our knowledge, this was the first work that attempted to enable online agent-side learning of argumentation strategies.
In this project, we applied hierarchical reinforcement learning techniques to symbolic agent communication languages to enable agents to learn time-extended policies in a strategic communication environment. (Fischer, Rovatsos and Weiss 2005) and (Fischer and Rovatsos 2005) are the main references for this work. Later, we also applied this framework to argumentation-based negotiation (Rovatsos et al 2005).
In this project, we applied hierarchical reinforcement learning techniques to symbolic agent communication languages to enable agents to learn time-extended policies in a strategic communication environment. (Fischer, Rovatsos and Weiss 2005) and (Fischer and Rovatsos 2005) are the main references for this work. Later, we also applied this framework to argumentation-based negotiation (Rovatsos et al 2005); to our knowledge, this was the first work that attempted to enable online agent-side learning of argumentation strategies.
This project is concerned with learning how to dynamically "stereotype" opponents in games with many opponents into different classes based on their strategies. We developed the AdHoc heuristic that was later on used by other researchers and applied to such domains as robotic soccer. Relevant publications are (Rovatsos and Wolf 2002) and (Rovatsos, Weiss, and Wolf 2003). We also developed methods for improving player satisfaction in video games based on classifying human opponents according to their level of skill at the game and adapting to this level (Collins and Rovatsos, 2006).
This project was concerned with learning how to dynamically "stereotype" opponents in games with many opponents into different classes based on their strategies. We developed the AdHoc heuristic that was later on used by other researchers and applied to such domains as robotic soccer. Relevant publications are (Rovatsos and Wolf 2002) and (Rovatsos, Weiss, and Wolf 2003). We also developed methods for improving player satisfaction in video games based on classifying human opponents according to their level of skill at the game and adapting to this level (Collins and Rovatsos, 2006).
In this work we developed a layered learning architecture capable of learning to play repeated games with large action spaces and greater numbers of players than in many other game-theoretic studies. The architecture consisted of a neural layer that would learn the utility function of the game, a genetic algorithm that would attempt to approximate the other's best-response strategies, and a layer using Probabilistic Ordering Models (a technique proposed for the architecture) that tries to model opponents' preferences. See (Rovatsos 1999), (Rovatsos and Lind 1999), and (Rovatsos and Lind 2000).
In this work we developed a layered learning architecture capable of learning to play repeated games with large action spaces and greater numbers of players than in many other game-theoretic studies. The architecture consisted of a neural layer that would learn the utility function of the game, a genetic algorithm that would attempt to approximate the other's best-response strategies, and a layer using Probabilistic Ordering Models (a technique proposed for the architecture) that tries to model opponents' preferences. See (Rovatsos and Lind 2000) for a summery of the main results.
Learning dialogue strategies (2000-2005)
Michael Rovatsos, Felix Fischer, Iyad Rahwan, Liz Sonenberg, Gerhard Weiss
In this project, we applied hierarchical reinforcement learning techniques to symbolic agent communication languages to enable agents to learn time-extended policies in a strategic communication environment. (Fischer, Rovatsos and Weiss 2005) and (Fischer and Rovatsos 2005) are the main references for this work. Later, we also applied this framework to argumentation-based negotiation (Rovatsos et al 2005).
Opponent classification (2000-2003)
Michael Rovatsos, Marco Wolf, Gerhard Weiss
This project is concerned with learning how to dynamically "stereotype" opponents in games with many opponents into different classes based on their strategies. We developed the AdHoc heuristic that was later on used by other researchers and applied to such domains as robotic soccer. Relevant publications are (Rovatsos and Wolf 2002) and (Rovatsos, Weiss, and Wolf 2003).
Opponent classification (2000-2007)
Michael Rovatsos, Marco Wolf, Gerhard Weiss, Brian Collins
This project is concerned with learning how to dynamically "stereotype" opponents in games with many opponents into different classes based on their strategies. We developed the AdHoc heuristic that was later on used by other researchers and applied to such domains as robotic soccer. Relevant publications are (Rovatsos and Wolf 2002) and (Rovatsos, Weiss, and Wolf 2003). We also developed methods for improving player satisfaction in video games based on classifying human opponents according to their level of skill at the game and adapting to this level (Collins and Rovatsos, 2006).
Layered learning (1999-2000)
Michael Rovatsos, Juergen Lind
In this work we developed a layered learning architecture capable of learning to play repeated games with large action spaces and greater numbers of players than in many other game-theoretic studies. The architecture consisted of a neural layer that would learn the utility function of the game, a genetic algorithm that would attempt to approximate the other's best-response strategies, and a layer using Probabilistic Ordering Models (a technique proposed for the architecture) that tries to model opponents' preferences. See (Rovatsos 1999), (Rovatsos and Lind 1999), and (Rovatsos and Lind 2000).
This project is concerned with learning how to dynamically "stereotype" opponents in games with many opponents into different classes based on their strategies. We developed the Ad Hoc? heuristic that was later on used by other researchers and applied to such domains as robotic soccer. Relevant publications are (Rovatsos and Wolf 2002) and (Rovatsos, Weiss, and Wolf 2003).
This project is concerned with learning how to dynamically "stereotype" opponents in games with many opponents into different classes based on their strategies. We developed the AdHoc heuristic that was later on used by other researchers and applied to such domains as robotic soccer. Relevant publications are (Rovatsos and Wolf 2002) and (Rovatsos, Weiss, and Wolf 2003).
http://www.cisa.inf.ed.ac.uk/agents/AgentsResearch.jpg
http://www.cisa.inf.ed.ac.uk/agents/AgentsResearch.jpg
Opponent classification (2000-2003)
Michael Rovatsos, Marco Wolf, Gerhard Weiss
This project is concerned with learning how to dynamically "stereotype" opponents in games with many opponents into different classes based on their strategies. We developed the Ad Hoc? heuristic that was later on used by other researchers and applied to such domains as robotic soccer. Relevant publications are (Rovatsos and Wolf 2002) and (Rovatsos, Weiss, and Wolf 2003).
We work on various topics in agents and multiagent systems. The following visualisation shows them along two critical dimensions, cooperative vs. adversarial and single-agent vs. multi-agent. Click on this map to find out more about the individual projects.
We work on various topics in agents and multiagent systems, mostly related to reasoning about interaction, whether this means modelling other agents, trying to resolve conflicts among interacting agents or learning what governs other's behaviour.
The following visualisation shows them along two critical dimensions, cooperative vs. adversarial and single-agent vs. multi-agent. As much as possible, we've attempted to represent similarity between different themes by proximity (although it is not always possible in 2D). Click on this map to find out more about the individual projects (feature not available yet).
Agents Research?.jpg
http://www.cisa.inf.ed.ac.uk/agents/AgentsResearch.jpg
Agents Research?.pdf
Agents Research?.jpg
http://www.cisa.inf.ed.ac.uk/agents/AgentsResearch.pdf
Agents Research?.pdf
http://www.cisa.inf.ed.ac.uk/agents/AgentsResearch.pdf
http://www.cisa.inf.ed.ac.uk/agents/AgentsResearch.pdf
Find out more about our projects soon.
We work on various topics in agents and multiagent systems. The following visualisation shows them along two critical dimensions, cooperative vs. adversarial and single-agent vs. multi-agent. Click on this map to find out more about the individual projects.
http://www.cisa.inf.ed.ac.uk/agents/AgentsResearch.pdf
Find out more about our projects soon.