 I'n falch i'r ffarnio, ond mae hwn wedi gwneud hynny o ffarnio ffarnio deisgrwm, a ddwy cael ei ddweud yma sy'n gweld yn effeithio surat diwylliannol y ffarnio yw, eich bod gweithio eich Llyfr. I'n ddweud am ymdannu ... yn ddod i'w ddweud yma eu du yn gwneud â'r systemu socialu. A'r ddweud i'r perthyniadau ffarnio deisgwyr o gyfan yw y ddweud, felly mae'r model agent wedi cael ei wneud yn bwysig o'r evaluatio a'r bwysig o'r ffordd. Felly mae'r simulatio, mae'r simulatio, mae'r simulatio sy'n ei wneud yn gweithio, mae'r idea sy'n ei wneud yn bwysig o'r programa'r model a'r bwysig o'r ffordd yn gweithio. Mae'r programa'r computer? Mae'n Gweith yma, mae'n dweud am gweithio gan gweithio gwyb surprise, mae'n gweithio yn ffwrdd â gyntaf unsi掉fodol, mae'n gweithio gweithio ond mae'n gweithio cyfle o'r mathymau, yn fwy fawr o yourf 어디w yw pob gweithio gyflym i'w adrodd o gennymol o agent o'r peiryaeth. Mae'n gweithio, mae'n gweithio, mae'n ddigwyd yn y gwybodaeth yn y modd yw… Gweithio, mae'n gweithio, mae'n ddigwyd yn gweithio. Mae'n ddigwyd yn ni'n ddigwyd yn ni, ac mae'n mynd i'n ddylch gyda'r ffaith, mae'n ddigwyd yn ddigwyd yn crynidol yn ei ddweud. And perhaps most importantly for our purpose, once you've got a computer model, a social simulation then you can run it as many times as you like you can experiment with it. Experiments we all know what an experiment is. In fact Nancy Cartwright spend some time talking about that kind of thing. Dewch ar ddefnyddio'r gwrthiaeth, ar ddefnyddio'r grwp어�au, a gweld i'r ddaeth a gweldiol i ddefnyddio'r gwrthiaeth. Rwy'n credu wasloedd eraill hynny yna allan i gychwynol â'r ddefnyddio'r ddefnyddio am yr wylan. ac mae'r ddweud o'r ddau, hynny'n dwi'n gobeithio, ac yn ddau'i cyfnod yn ddweud. Yn ddweud, mae'n ddweud i gweithio'r swyddion cyflogol, o'r cyffredinol, o'r cyffredinol cyffredinol, mae'n ddweud yn ymryd o'r gwyffr o'r cyffredinol, Onw'r ysgolch oedd yn candliadau, oedd yn mynd fathafol, oedd y teimlo ond argynno fyddio'r cynnwys bethau mae'r dyn nhw i'ch gynfyddyddiau yn gofynu iawn, lle gynnawn ei rgodio'r hunain. Yn ymwysil yma, yw'r unrhyw yma, sy'n cael ei wneud bod y rhywbeth yma yn gwneud fathafol â'r gweld yn yr yw. mae amser i na fydda diwethaf. Mae'r local yma'n gweld gyda hwnnw, y dications i gyda gweithio gyda'r dweud o'r cyfnodol, a phedda'r cymddiad a facesg wedi cyfeirio dim o'r cyfrwysol byddai gyda'r model yn ei gwleidydd. Fyn ni'n meddwl y cwm yn symuol. ond we build a model and we try the application of a policy on the model instead of the real world. So if you want to find out, as a simple example, what the consequences of raising income tax at the lowest level or applying universal credit or something of that kind one way would be to take a random sample of people and give them the new tax. But there are very obvious reasons why one can't do that. But what one can do is build a computer model of a sample, a random sample, even of individuals and see what happens to their income or their net income as a result of tax and benefit changes. And that's been done for many, many years. So we do experiments on the model and if the model is a good one and that's a big if, it will react in the same way as the target, the real population would have done. And we can repeat the experiment many times and that in that way we can try different policies, we can average out over random variations. OK, so building a computer model is one thing, but what do I really mean by a computer model? I'm actually talking about a specific sort of computer model, one we call an agent-based model. And here's an illustration of what an agent-based model is and we start off with nothing. And from the nothing we grow and environment, that's just a computer representation, perhaps of a geography, a space, perhaps of a network, whatever seems appropriate. Into that environment we put computer agents representing the actors in the real world. But they may be representing individual people, they might be representing organisations, even nation-states. We program them to interact with each other and that's a really important part and we're talking about the social world, so we need to represent the social interactions and we give each of our agents a degree of autonomy. By that I mean that what we do is we say that these agents behave according to the context in which they find themselves. So in this trivial example I've just got our agents to follow their nearest neighbour and we end up by them all sort of joining in to a big crowd in the middle. Of course a real model we would have much more sophisticated, much more complicated than that. But there we have what an agent-based model is. We have agents behaving in an environment over time. Here's a couple of examples to make that a bit more realistic. If we wanted to model the spread of bird flu or any kind of epidemic, the standard way of doing that is to use some mathematical equations. It's called the SIR framework in which what we do is we calculate the number of people who are susceptible, the number of people who have been infected, the number of people who have been removed, that might be either because they were covered or because they died, and in some diseases we also have an intermediate stage where they've been infected but haven't actually become infectious yet. That's just a piece of rather simple mathematics in epidemiology, but if we convert this into an agent-based model where we have individual agents, the kind that I've just been talking about who get infected when they bump into other agents that are infected, we can model this process. Now why would we want to bother about doing that if we can do it all with a few equations? Well the thing is that if we have an agent-based model we can build in some interesting variations. For example, we can give our agents the sort of stereotypical behaviours like going to work and meeting their work colleagues and getting sneezed over at their desk, coming home and infecting their family and so on and so forth. So we can build into our agent-based model the routines of daily life and thus make both the model more realistic and also begin to start indicating what the most appropriate of good ways of vaccinating the population against the disease. So we get flu epidemic graphs by measuring our model, by putting probes into our computer model, looking at how many people are infected, are dying or whatever. Okay, so we can use the model to study the effect of various different policies, various different parameters, various different distributions and test different control strategies. Here's another example of a model which was designed to look at the English housing market and what we have here is the agents are households which are trying to buy a house or sell their house and the square in the middle represents a town and we can instrument this model with all the different things like the house prices as they go up and down over time. The important thing I want to emphasise about this, I won't go into any detail at all, first of all we can represent the actual process of buying and selling a house from the point where you want to sell a house, you go to the estate agent, you ask for valuations, you wait around for people to become interested in the house, you take them around the house, they have to go and try and find a mortgage and all the rest of the processes involved. Every step of that kind of process can be represented in the model. The second thing that's important about this is that these are dynamic graphs here that these agent based models work through time unlike mathematical equations. Just to summarise agents are units that have behaviour, they act with a simulated environment and we look at the macro level or emergent characteristics in our computer model. Now I don't perhaps need to say a lot about evaluation but here's the official word about what evaluation is. It's taken from something called the magenta book which is the Majesty's Treasury guide as to how policy people in government should do evaluation and that's their definition of evaluation. The bits to emphasise are the way that it examines the implementation and impacts of a policy to see whether the anticipated effects were actually realised. The problem with evaluation is knowing what would have happened had the policy not been implemented. Despite what Nancy Gartwright's dismissal of counterfactuals is saying actually people really do want to know has my policy had an effect. This is not for scientific reasons if you like it's because well they have to justify their policy and when you justify a policy it is implicitly or explicitly against the business as usual the counterfactual of no policy. But establishing the counterfactual what would have happened is not simple and it's especially not simple when you have the kind of complexity that was being shown in the previous talk. Here's just a very quick summary of some of the reasons why it's not so simple. First of all complex systems are intrinsically difficult or impossible to make point predictions about. The reason why I have this backdrop on this slide is because the weather is another complex system and we know that even with supercomputers we can't predict the exact weather that's going to happen in more than 10 days ahead. An example of the fact that you can't predict many aspects I'm going to be oversimplified here many aspects of a complex system although you can, well I won't go into the details. The second area where it's really difficult is you get these things called tipping points. Like this example because it's so clear this is a graph of the price of solar panels the sort of things you can put on your house roof over time and this is the blue shows you the number of installations of solar panels and what you do not see is as the price goes down the number of installations goes up. There is a point here where you get a tipping point where suddenly you get a growth in installations. It takes time for that for solar panels to become something that you do install and just think about if you were trying to evaluate a policy to subsidize solar panels notes that the time at which you did the evaluation will be extremely critical. It goes back to what Emma was saying about time, the importance of time. And thirdly and lastly for the moment I wanted to mention the idea of path dependence. This is a model about again going back to bioenergy as we've talked about in the previous session. This is a model of the installation of anaerobic digesters which are things which take for example food waste and turn them into gas and energy. The graph is about the effect of financial incentives of subsidies in building these digesters. What is it all about is that as the financial incentives increase the chances of a number of anaerobic digesters that are installed increases as you might expect it becomes more and more profitable to run such things but coming down again if the government then decreases the subsidies these anaerobic digesters don't all immediately start going out of business because they've already been set up. So exactly the path of financial subsidies is very important. If this government stopped there then end up with no anaerobic digesters but if they went up to the top and then down to four they end up with a lot of digesters. So path dependence is another area where straightforward linear frameworks for evaluation are not going to do the job. I'm actually going to because we're really out of time I'm going to skip that and say that the combination to deal with these kinds of problems one way of tackling them is to build an agent-based model and to use that because any agent-based models can cope with the kinds of complexity kinds of part dependence and tipping points and so on that I've just been talking about because to put it very simply they are trying to represent what's actually happening in the world if there's a tipping point in the world you should be able to model that in an agent-based model just as much as it happens in the real world and the advantage of doing that is that you can then run that model many times and explore the possibilities of having different policies or how even with the same policy the effects may vary as the other things change. However, just to say that if and again I want to emphasise it with a big if if the model is a good one that is if we've got the right simplifications the right boundaries around the system and we've modelled the right processes within the agent-based model then we ought to be able to learn a lot about the system and about the evaluation that we're interested in. However, life is never simple. All useful ABMs are stochastic which means that they've got random elements in them so that means that we need to run the model many times and take averages even taking averages may not be a sensible thing to do if the distribution of results is not far from normal but what we then want to do is to compare runs with the policy and without the policy that is the counterfact, compare the policy with the counterfactual but we have to compare distributions not points and if the distributions are sufficiently different then we can conclude that the policy has had some effect at least in our virtual world. I've managed to do that in the time and we've got all of two minutes left so I'm just going to finish off by telling you about where you might go if you want to explore these ideas further. There is an open access journal called The Jazz the journal of artificial societies and social simulation if you Google jazz with three S's and that's got lots of stuff about the application of agent-based modelling to aspects of social policy and politics from authors all over the world and there's the link that you can use and you can also look at the CCAN website which includes things we call the evaluation policy and practice notes which are briefing papers about a whole range of evaluation methods including user agent based modelling or you join us on Twitter, Facebook or LinkedIn that means that we have 30 seconds for a question.