Nowadays, social organizations (at macro-level) can be represented as complex self-organizing sys-tems that emerge from the interaction of complicated social behaviours (at micro-level). Modern multi-agent systems can be employed to explore "artificial societies" by reproducing complicated social behaviours. Unfortunately, promoting interactions only among pre-set behavioural models may limit the capability to explore all possible evolution patterns. To tackle this issue, we aim at discovering emergent social behaviours through simulation, allowing human people to participate in the simulation environment, so that the range of possible behaviours is not pre-determined. In order to support this new approach, we propose a system architecture that is able to support an endless session level between a software agent and a human player (called participatory framework). In par-ticular, while network faults or human low reactivity do not allow the human being to control his agent, this system architecture adopts a virtual player mechanism (called ghost player) that takes control of the agent driven by the user. The advanced version of such a ghost player relies on sub-symbolic Machine Learning techniques for mimicking the strategy of the off-line human being.