Description

Flatland is a prototype two-dimensional (2-D) multi-agent world inspired primarily by the virtual worlds met in computer games and by biological aspects like animal foraging and mate-seeking. The main purpose of this simulated world is to be used as a test bed environment for investigating evolutionary and gradient-based (to a lesser degree) learning techniques and furthermore, their ability to generate cooperative obstacle-avoidance and target-achievement behaviors.

Flatland is a complex environment in that its properties depend on the interactions between agents—moreover, the dynamics of agent behavior depend on discontinuously changing relationships between the agents as they prove their double task. In that sense, complexity in Flatland is a feature determined by the number of agents present. 

Further interested in FlatLand research?

Papers about FlatLand

• G. N. Yannakakis, J. Levine, and J. Hallam, "Emerging Cooperation with Minimal Effort. Rewarding over Mimicking,'' IEEE Transactions on Evolutionary Computation, vol. 11, issue 3, pp. 382-396, June 2007. [pdf]
• G. N. Yannakakis, J. Hallam and J. Levine, ''Evolutionary Computation Variants for Cooperative Spatial Coordination,'' in Proceedings of the 2005 IEEE Congress on Evolutionary Computation, pp. 2715-2722, Edinburgh, UK, 2005. [pdf] 
• G. N. Yannakakis, J. Levine, J. Hallam, and M. Papageorgiou, "Performance, robustness and effort cost comparison of machine learning mechanisms in FlatLand,'' in Proceedings of the 11th Mediterranean Conference on Control and Automation MED'03. IEEE, June 2003. [pdf]

FlatLand Simulations available for download

Instructions: Download and unzip the files. Run MAGIA_GR.exe

• 20 Animals (agents controlled by Artificial Potential Fields)
• 20 Humans trained with Back-Propagation
• 20 Humans trained with the GGA algorithm
• 20 Target Achievers

N.B. Windows Vista do not support fullscreen mode for DOS-applications. Go through these guidelines if you want to view FlatLand simulations in Windows Vista.