We study certain types of Cellular Automata (CA) viewed as an abstraction of large-scale Multi-Agent Systems (MAS). We argue that the classical CA model needs to be modified in several important respects, in order to become a relevant and sufficiently general model for the large-scale MAS, and so that thus generalized model can capture many important MAS properties at the level of agent ensembles and their long-term collective behavior patterns. We specifically focus on the issue of inter-agent communication in CA, and propose sequential cellular automata (SCA) as the first step, and genuinely Asynchronous Cellular Automata (ACA) as the ultimate deterministic CA-based abstract models for large-scale MAS made of simple reactive agents. We first formulate deterministic and non-deterministic versions of sequential CA, and then summarize some interesting configuration space properties (i.e., possible behaviors) of a restricted class of sequential CA. In particular, we compare and contrast those properties of sequential CA with the corresponding properties of the classical (that is, parallel and perfectly synchronous) CA with the same restricted class of update rules. We analytically demonstrate failure of the studied sequential CA models to simulate all possible behaviors of perfectly synchronous parallel CA, even for a very restricted class of non-linear totalistic node update rules. The lesson learned is that the interleaving semantics of concurrency, when applied to sequential CA, is not refined enough to adequately capture the perfect synchrony of parallel CA updates. Last but not least, we outline what would be an appropriate CA-like abstraction for large-scale distributed computing insofar as the inter-agent communication model is concerned, and in that context we propose genuinely asynchronous CA.
