Automatic Recognition of Collective Emergent Behaviors Using Behavioral Metrics

Collective emergent behaviours are commonly seen in nature such as in flocks of birds and schools of fish. These behaviours are the results of years of evolution and have been studied in artificial agent systems in a wide range of application areas such as robotics, serious games, and crowd simulations. Automatic recognition of such collective behaviours is imperative in such application areas in order to measure and improve the effectiveness and efficiency of the artificial agent systems, especially when it involves machine learning approaches where human labelling is not feasible. While it is easy for the human eye to recognise collective behaviours, this is an extremely challenging task for a machine to automatically recognise them as such emergent behaviours cannot be captured by a simple mathematical equation. This paper investigates how emergent behaviours can be automatically recognised through capturing the behavioural aspects of the collective nature of the agents’ performance. We identify seven metrics such as grouping, order, and flock density that can capture diverse and distinct emergent characteristics of agent behaviours. Five machine learning models that use a combination of these metrics as features of a range of representative behaviours were trained to investigate the potential of automatic recognition of collective emergent behaviours. The evaluation results show that training the machine learning models with the proposed approach enables automatic recognition of a range of diverse emergent collective behaviours. Further, we conducted leave-one-behaviour-out experiments on the representative behaviours and the metrics used. The results confirmed that each behaviour and metric have a unique impact on accurate recognition of emergent behaviours in collective agent systems.