Dynamic Agent Compression

We introduce a new method for processing agents in agent-based models that significantly improves the efficiency of certain models. Dynamic Agent Compression allows agents to shift in and out of a compressed state based on their changing levels of heterogeneity. Sets of homogeneous agents are stored in compact bins, making the model more efficient in its use of memory and computational cycles. Modelers can use this increased efficiency to speed up the execution times, to conserve memory, or to scale up the complexity or number of agents in their simulations. We describe in detail an implementation of Dynamic Agent Compression that is lossless, i.e., no model detail is discarded during the compression process. We also contrast lossless compression to lossy compression, which promises greater efficiency gains yet may introduce artifacts in model behavior. The advantages outweigh the overhead of Dynamic Agent Compression in models where agents are unevenly heterogeneous — where a set of highly heterogeneous agents are intermixed with numerous other agents that fall into broad internally homogeneous categories. Dynamic Agent Compression is not appropriate in models with few, exclusively complex, agents.