Hi everyone, I'm Emanuele, I'm the CEO and co -founder of Wizen. In Wizen we are developing a decision intelligence platform where we develop custom AI models for optimization problems and decision making problems.
So our idea is to develop a platform that can enable people or customers or organization companies to develop their own AI models to solve their own problems without involving an external AI expert or specialist.
So our idea is to give them these instruments to use their own data and their technical knowledge about their problem to make their own model alone, without involving someone like us to in a consultant way so the first step in this process is to build the model to
solve your problem and our idea is to obtain such a model in this case only with a chat based interface but usually you have also the data so data from your machines from your historical historical domain and mix that data with the knowledge from your experts so let's say that in this case we don't have any data
I just pick a random challenge or posted by a company in this case a railway company and since I don't have their data they I don't have access to their their data I asked the system to generate a toy model that can maybe I can present to them as an example of what we can do for them so let's say
that I give the interface like three phrase of context and constraints on how to model the problem so in this case I just report that this is a challenge to To create a simple causal model with around 10 variables involved, 10 factors, that can be extracted from the challenge itself in the text below.
To keep the variables as simple as possible, so for example, if the challenge report the weather conditions, we are not going to use the weather as a whole variable, but to split that variable into the temperature and raining condition maybe, the humidity. so to keep the model basic and then I just report the challenge the title
challenge and the text that in this case is just two phrases that is to develop a solution AI solution that combine real -time operational data with a list of few variables that that they are considering on day or that the company believe is going to be useful to assess what they want to measure that in this
case is to improve the perceived pontuality of the trains and reducing the disruptions in the system.
So we run this very brief chat and the model itself, so now let me resize a bit.
We'll extract this model from the chat alone if we also upload the data so a structured data or unstructured data like maybe some the topology in this case the topology of the network the training system the manual of some system subsystem in the network and so on the system is also going to use that information to build the model in this
case I don't obviously have access to this so it's just the representation presentation extracted from the the chart.
Obviously then the agent also tried to explain you what he did, so the variables that he identified as relevant and how they are connected among each other, so which are the relation among the variables.
So for example, let's pick some as an example, there are exogenous variables that are the variables without any cause In the model that is the temperature and the rainfall so the weather
He choose to include the urban events So if there is a big event in that city and that you are trying to model Obviously this is going to impact the railway network the number of stuff involved in that period and the feelings on the train itself and
The point of the model is to starting from this variable variable, understand which variables downstream are affected if we change something in these ones.
So for example, in this case, let's say that the easiest one is to see this one. So from the urban events, if you have a big event, then the model understands that the number of passengers is going to increase in that moment, and this is going to make that this can may impact on the punctuality of the trainings so this is the model that we can develop in this case just by putting something then
let's say that we have access to the expert knowledge in the company in this case I make the role of the expert so I know that I have some knowledge about the problem and i know that some other variables they may affect the the problem and so i asked
the system to add some variables so for example the initial delay of the train if i'm trying to understand the delay at some point in the in the travel of the train the initial state so the initial delay is going to count a lot so i asked the system to add that the trainer availability availability and the recovery time so in the train network they usually add some minutes between one session and the others as bonus to recover from previous delays
and since I know that these are important factors in the problem I asked the model to add them so this is the final model with also these other variables and we can say that for example the initial delay is going to to affect directly the train delay later on and this is a very simple cause and
effect relation because it's going to directly reduce the number of means of delay because we are using that extra bonus that we give directly directly into the network.
So once we have such a model, we can also use that to explore, for example, the model.
So I can ask the system to explain to me some part of the model. So, for example, we can ask which are the main causes of the terrain delays represented in this model, and how they influence each other.
and in this case also which factors have the greatest impact in the final delay that is the main outcome that we are interested in so in this case the system find that the direct causes are the initial delay obviously the dwell time that is also affected by the number of people that is also affected by the number of the if there is a big event in the city then you find all the causes in the factors that we are interested in.
So let's say then we can also simulate something. So now we have let's say that we want to assess what is going to happen to our network if we have a particular condition.
So for example let's say that it's a rainy day and there are several events in the area and this is is going to cause an increased crowding in the stations. And some of our trains have failures. So how this is going to affect the delay.
So here we can see, in this case, the explanation. So in this specific scenario, so the rainfall is going to affect the train signal. so it's going to put a problem into the network itself, this is going to cause micro disruptions in the network and in the schedule of the trains and this is going to delay the trains.
Then obviously we have another path from the urban events that is going to increase the number of passengers, increase the time that the passengers need to board on the train in all the station is going also to increase the delay and finally if we have obviously a fault on the
trains this is also going to cause disruption and delay again the train so this is in this case we have three factors that we ask them the system to take into account in this in a specific condition that are going to impact the final delay.
So now in this case we don't have data to model the specific form of the function, so how much a factor is going to impact the final outcome.
But in the case we have the data from the company, we can also use this data to find how much each of this factor is going to impact the final delay and this is particularly helpful in the case where we
want to intervene on the system so for example we want to compare two different strategies to make to manage the disruption maybe one is to simply increase in the recovery time so we give the trains one one extra minutes between two stops or we can improve the number of things that are
available at each time obviously this is going to be the first one is going to be very easy to do even if we have to reschedule maybe a lot of other trains on the network the second one is more effective but it's expensive to do so the system can compare to different strategies compare the the expected outcomes from both the strategies,
and let's say you can estimate the expected impact and also the expected cost of implementing both strategies so that the decision maker in the company can decide which one is going to be the best solution.
so obviously then is also try to merge the two solutions we can also try to combine different kind of strategies so we can ask the system to combine the strategies to combine the effects to see if the two single effect can be summed up or maybe one strategy is going to impact negatively impact the other one
So, in this case, the systems, since we asked for the increased recovery time and the number of available trains, say that there is no overlap because they are onto different cause and effect relation paths, so we can get the best of the two channels without, let's say I will use the same words, that we can expect that the combining strategy is to be more effective than one, so the one or the other alone, because we do not expect that using both at the same time is going to influence the other one.
So that's it, basically. And it's, that's it.