Abstract
A requirement for collective action is increasingly encountered in open multi- agent systems. However, in open systems, with no central controller and in- dividual autonomy, and with possibly inconsistent beliefs and conflicting goals, achieving successful collective action typically requires some form of self-organisation rather than relying on luck, magic or Delphic intervention.
Theories of collective action from psychology, economics and political science have shown that successful collective action can be produced by coordination mechanisms such as alignment cascades or common knowledge. Therefore, in this thesis we aim to develop a computational framework for collective action in multi-agent systems in which agents can represent and reason about such coordination mechanisms to inform their decision-making or action selection.
Inspired by social scientist Obers’ work, who argued that the success of classical Athens was the result of its shared commitments, social values and specific pro- cedural rules, we consider two of the knowledge alignment mechanisms such as alignment cascades and common knowledge that were used in classical Athens in order to allow citizens to join together for a common good. Therefore, the aim of the thesis is to develop a computational framework for collective action in which agents can represent and reason about such coordination mechanisms for decision-making or action selection.
To investigate how alignment cascades contribute to collective action we im- plement a Bayesian model of an alignment cascade where agents make decisions in a sequential fashion based on private information and observation of other agents. Simulation results show how this Bayesian model reproduces and in practice outperforms another probabilistic model, by suppressing “incorrect” cascades.
As a second step, we choose another knowledge alignment technique: the use of common knowledge and the existence of a shared cooperative ethos to foster collective action. We propose a probabilistic logical model of this reasoning problem based on Markov Logic, and analyse a scenario of collective action based on Ober’s analysis of an historical trial in classical Athens.
In our Markov Logic model we directly encode the existence of common knowl- edge to study how this common knowledge can emerge in practice, we revisit Lewis’s analysis of common knowledge, widely considered to be the first such work. Unlike the standard approaches to logical modelling of common knowl- edge, Lewis defined properties that are sufficient for information about the world and other agent’s reasoning mechanisms to lead to chains of iterated mutual knowledge, without directly reasoning in those terms. We propose a practical interpretation of Lewis’s approach in terms of a situated agent with a reasoning mechanism based on a forward-chaining rule engine and ‘theory of mind’ rules allowing other agents’ beliefs and rules to be modelled.
Following that, we combine all these models in simulation of this scenario to show how common knowledge can emerge among the agents and how align- ment cascades contribute to the collective action. The significance of this work is that to combine these aspects in simulation with reasoning will allow us to show how knowledge alignment mechanisms and joint action can be realised in a computational agent.
In conclusion, formal studies in logic and philosophy have shown how theories of common knowledge can be highly expressive and predictive. The results of our simulation show that it is possible to extract a model for a subset of a social science theory that is computationally tractable but also philosophically and psychologically plausible.