Knowledge stored in a KnowledgeNet is combined to show predictable features of several variables of interest. It shows interesting features of the world that have been always true until now. They serve as the basis in multi-goal multi-decision variable decision tasks.
Our learning systems (combination of a databases a logic module and meta-laws) search for objects that have properties, which are emergent relevant for the decision, and combine them into a Worldview (Conception of the World), that can be used for prediction and decision purposes. As an example, we choose the loan portfolio of the peer-to-peer lending company Lending Club (122,444 loans).
We have searched for objects in the first 100,000 loans. We can find a set of objects for the loans on the marketplace operated by Lending Club, to which until now multiple always-true statements about the above-mentioned properties of loans were attached. The result is shown in the below movie. The movie shows an up to this point always-true Worldview defined by the properties of a credit portfolio, which are relevant for a decision; and a prediction horizon of TP=15,000 loans.
Investors can find dominant strategies to select loans. All objects in the top right corner always had a higher return than the average and a higher return than the average of the respective rating class. (This result shows recurring “errors” in the pricing-system of lending club.) The point in the respective cuboid represents the current state of the objects for all depicted dimensions.
If the movie is started and as time passes, the current state (point in the respective cuboid) is mostly located inside the correct and therefore future expected intervals. It can also be seen that sometimes the learning system “adapts” its conception of the world. For some properties this behaviour was allowed. But for every point in time the movie shows the current state of decision relevant properties of Lending Club’s database that was always true. No object disappears over time. This will happen, if one of an object’s important properties is falsified by new observations. We also see that the Worldview is pretty stable. Especially the dominant strategies (top right corner) stay dominant all the time (compared to the ones in the bottom left and bottom right).
We should remark that the decisions made by our learning system contradict often the currently used decision rules derived from decisions under risk. Nonetheless it would have been always profitable to follow the decisions made on the basis of T-Dominance.
This phenomenon can be observed for many decision problems. At the moment there are “free lunches” everywhere.
If the prediction horizon is increased to TP=22,500 loans, the resulting Worldview contains more objects (36), some of which (3 objects) are indeed falsified and therefore disappear from the Worldview.
The Worldview, which is created by the learning system, differs from conventional modells in the following points:
- The Worldview consist of many single laws, that where always true until now. This approach safely prevents overfitting.
- Not necessarily every measurement belongs to a set of measurements, for which relevant properties can be predicted.
- Properties of an object relevant for decisions naturally allow the pursuit of multiple objectives.
- In the next step it will be possible to reproduce movement and acceleration of Objects.
- An object “Me” can be integrated into the conception of the world. This object will show the resulting properties of the application of decision heuristics by the learning system.
- Ultimately it will be possible, that the interactions between the “Me” and other objects in the Worldview are observed by an encompassing-system. This creates a learning system that is able to look at the interaction of a representation of itself with objects in a world – consciousness.