Cellular Automata, Agent-Based Modelling and Simulation

From COSNet

Since the release of Conway’s Game of Life, cellular automata have been used as models in many areas of the physical and spatial sciences, biology, mathematics and computer science, as well as in the social sciences. They are a very useful modelling platform, as cells on a grid that switch on or off according to states of neighbouring cells can represent a host of dynamic phenomena – individuals, attitudes or actions, for example. A cellular automaton (CA) models any world in which space can be represented as a uniform grid, time advances by steps, and the “laws” of that world are represented by a uniform set of rules which compute each cell’s state from its own previous state and those of its nearby neighbours.

Once we wish to develop automata that are somewhat more complex in their internal processing and consequently in their behaviour, we enter a different world known as agent-based modelling or simulation (or even multi-agent systems). Such automata are conventionally called agents, and there are several growing streams of thought on how the agents can be designed, built and used. While there is no generally agreed definition of what an agent is, the term usually implies an autonomous, intelligent entity that may interact or communicate with other autonomous, intelligent entities. As with CA, there are rules governing interactive behaviour and the agents “operate” in or on an environment of some sort.

The agents emanating from the literature on (distributed) artificial intelligence often correspond to self-contained software and/or hardware (e.g. robots) that control their own actions based on their perceptions of their operating environments. Multiple agents are designed to work together to achieve a desired goal. Typically, their goal is pre-specified and they are engineered and controlled to achieve it with very little tolerance for error. Although purposive, systems in which human agents interact with ecological systems, for example, display open-ended outcomes. Here the collective behaviour is unknown in advance, but emerges during the simulation. Some emergent outcomes may be unexpected and undesirable. Artificial life has been the source of inspiration for this open-ended kind of simulation.

In agent-based simulation, rules governing agents’ behaviours can range from simple “IF—THEN” clauses to quite sophisticated machine learning algorithms (such as genetic algorithms) that allow agents to modify and improve their behaviour during the simulation. Data mining and visualisation are used either to ensure that agents behave in ways that realistically depict how individual decisions are made in that system, or to extract knowledge to understand the dynamics of the problem and its behaviour.Parameters of the model are set to represent a situation of interest and the model is run for several hundreds of iterations, until a preferred solution is found. Where possible, simulation models are calibrated against historical data to ensure that the model is accurately replicating the behaviour of the real system.

Agent-based simulations can provide valuable information about the dynamics of the real world(s) that they emulate. Complex systems scientists see them as more realistic than equations-based methods alone, because they are built synthetically “from the bottom up”. As a wide variety of agents interact within a social science simulation, for example, it shows how their collective behaviours govern the performance of the entire system – for instance, the emergence of a successful product, a congested area of traffic, or a polluted water catchment. This is of great benefit to key stakeholders, because they see a role for each of them (as agents) in the simulation itself, as well as an opportunity to learn from the simulated outcomes. Thus an area of the field that is growing rapidly is the use of participatory agent-based simulation to assist with human – environmental interactions (e.g. natural resource management).

Agent-based simulations can also capture reality more effectively, without assumptions such as monotonicity. Different type of noise, behaviours, and/or effects can be incorporated into the simulation as needed.

Agent-based simulations are also powerful tools for “What if” scenario analysis. As certain agents’ characteristics or behavioural rules change, the impact of the change can be seen in the model’s collective output. It is these adaptive learning features that give simulation an edge over more traditional modelling and optimisation methods. Perhaps most important of all, the computer can generate outcomes or strategies that a scientist or stakeholder might never have imagined.

The COSNet program in this area will be integrated with the CSIRO CCSS Agent-Based Modelling working group