Skip to 0 minutes and 10 secondsThe last years there's an increase in the interest in using psychological theory to formalise these agent-based models and to compose more realistic agents. In my own work, I used a framework called consummate to simulate consumers. And two critical things can be seen in this approach. First, that the simulated agents, simulated consumers have different needs-- a need for subsistence, a social need, and a personal taste need. And very often, these needs can be in conflict. What you personally like, might be not social approved. A second part in this model is the decision making of the agents.
Skip to 1 minute and 3 secondsAnd I make a distinction between four basic strategies, called optimising, which is basically trying to make the best choice possible, inquiring, where you ask experts about their opinion, what is used a lot on the internet nowadays. A third one would be imitation, where in a closed group you follow the behaviour of your friends and family. And the last strategy is repetition, which is basically addressing habitual behaviour. And of course, more than 90% of our behaviour is habitual, so it's important to incorporate this in the architecture of simulated agents. in a model you can see how this works for one agent and that you can have different critical components in a model.
Skip to 2 minutes and 0 secondsFor example, if you think about new behaviour, you will update your memory. You will learn. So learning is in a model. On the other hand, if your simple repeat behaviour in your inhabitual mode, you will not update your memory, which may become outdated. So your habit may be far from optimal.
Skip to 2 minutes and 27 secondsThe main challenge here is how can we parametrise such a model? What does it mean? We have a nice architecture, but, suppose, you want to model a farmer data selecting a particular crop or you want to model a car driver that's is facing the choice of a new car-- should it be an electric one or a fuel one? Very often from research we have a lot of data available. And of course, we can use these data to put it all in our simulation models. But keep in mind, we want to demonstrate the basic principles of emergence. So we have to keep our models relatively simple, because, otherwise, we do not understand what is happening, just as in the real world.
Skip to 3 minutes and 24 secondsSo we need a certain level of simplicity. So that's a very important step selecting what kind of data is really critical of getting the model as realistic as possible to demonstrate the process we're interested in. And for example, if you want to model the diffusion of electric cars, the social susceptibility, these norms appeared to be very important. In particular, the anti-conformist motive might be very important for some agents to decide purchasing an electric car to show off that they are different from the rest. It's shown that this works in reality, so we should capture that in a simulation model as well.
Skip to 4 minutes and 19 secondsNext, the question is if you have a model, are the outcomes realistic? Can we trust it? Does it capture reality? Now, very often, of course, we use these models to get an idea of possible futures. But the future is not today, so we cannot validate our model outcomes against a future that is not there yet. So we have to do it with the data that we have available today. But then we run into a very nasty problem. If you really understand that society is complex, you realise that today is only one manifestation of many possible states that could have happened. Today could have been really different.
Skip to 5 minutes and 12 secondsSo it's also very dangerous to validate your model against data that describe how the world is now, considering that the world could have been different. So what I usually propose is that the processes that grow in your model, and in particular the processes that grow in the individual behaviour of the agents, reflect what we know of behaviour of real people. I often call this life histories. So the life histories of real people should be reflected in the histories of the simulated agents.
Skip to 5 minutes and 57 secondsNow, if we are trusting our model in the sense that it does what you expect it to do, that it creates realistic behaviour, what can we do with a model? Well, classical prediction? That's no way. We know that we're dealing with complex systems and prediction is not a good thing. So what can we do then? Classical experiments? What happens if? If you realise the vast amount of possibilities, you can't conclude that there are too many what ifs. There's no limit to be number of experiments you can conduct. So what I propose to do is focus on gaming. And gaming has a way of playing around with the dynamics of the system that you have been modelling.
Skip to 6 minutes and 57 secondsAnd using this game you learn to become familiar with the dynamics of the system. Usually I compare this with a flight simulator. I feel very comfortable in a plane, knowing that a pilot has crashed numerous times. Fortunately, in a simulator, of course. But this is very good, because the pilot has learned to understand the system and has become very adaptive to signals that identify possible problems. And I think that this agent-based modelling technology can be used to develop social flight simulators, that we have a training tool to explore the dynamics of social systems that may help us in finding effective means of managing our own society.
Modeling social systems
This lecture explains some of the challenges and opportunities of modeling social systems. Agents could have different needs and decision making strategies. How do you make an Agent-Based Model based on such different needs and strategies?
© University of Groningen