<< content                              Chapter 6

Predicting the future in terms of sequences

In previous chapters we only talked about one gameplay for a football game.  In order to predict the entire 4 hours of the football game, many sequential gameplays are to be predicted.  FIG. 24 depicts one gameplay in the football game.  This prediction tree contains important hierarchical predicted models.  Each predicted model contains the strongest objects grouped together in terms of dependency and importance.  All objects of the gameplay are hierarchically structured for all objects involved (large objects like human beings or small objects like a blade of grass).

FIG. 24

Predicting sequential events

For simplicity purposes, a gameplay is represented with a G and gameplay1 is called G1.  FIG. 25 is a diagram depicting a sequence of gameplays for the football game (G1-G8).  Usually, the importance of a gameplay is based on the distance from the current state.  The closer the gameplay is from the current state, the more important it is.  For example, G1 is closer to the current state so it has a priority of 50%.  G2 is farther away from the current state so it has a priority of 30%.  The higher the gameplay’s priority percent, the more virtual characters are assigned to predict the gameplay.

FIG. 25

Predictions are done incrementally.  The virtual characters will predict gameplay1 and check to make sure the future possibilities are accurate.  Next, it will predict gameplay2 and combine that with the previous prediction (which is gameplay1).  Then it will predict gameplay3 and combine that with the previous predictions.  The idea behind a prediction tree for sequential events is to make a prediction sequence lengthier, but at the same time, to predict the “whole” sequence.  For example, if the prediction tree predicts gameplay4, it also has to consider its previous sequences (gameplay1-3).

The prediction tree is constructed incrementally as the virtual characters add longer sequences.  For example, if the virtual characters are predicting G1 and G2, the prediction tree will generate predicted models for G1 and G2 and all of its upper and lower levels.  Part of their future sequences might be a part of the prediction tree as well.  For example, G3 and P2 might be a part of the prediction tree.  The prediction tree is assuming that the virtual characters will be predicting G3 in the future.  As more sequences are added, the prediction tree will add more branches of predicted models. 

Each predicted model is interested in ranking their future possibilities.  At the U level of the prediction tree, they are interested in creating a general ranking of future possibilities of gameplay1-gameplay8.  At the P level, they are interested in creating a general ranking of future possibilities according to their neighbors.  For example, the predicted model P2 is interested in predicting future possibilities based on G2-G5.  At the G level, they are interested in creating a detailed ranking of future possibilities according to their neighbors.  For example, the predicted model G4 is interested in creating a detailed ranking of future possibilities based on G3-G5 and most of their lower levels (FIG. 25 and FIG. 26).

FIG. 26

The sequence that each predicted model in the G level is limited to is about 3 gameplays.  For example, G4 is responsible for the sequence G2-G5 and G6 is responsible for the sequence G5-G7.  They will create ranked possibilities for these limited sequences.  FIG. 27 is a diagram illustrating a predicted model that includes sequences.  There is a focused objects, peripheral objects and each event has a sequence length (an event is an object).  The ranking of future possible events is based on the sequence length and the focused objects involved. 

FIG. 27

To complicate things more, the sequence length for a predicted model can change it’s scope.  Also, very general predicted models like the P level don’t really have a fixed sequence length to work with because events can be fragmented.  For example, language can encapsulate entire events.  A sentence like “the cowboys win the game by a large margin” can encase the entire game.  Important fragmented events in the football game can be represented by sentences.  In the U level, the predicted model might highlight 4-5 important events in the game.  These important events are extracted from the entire gameplays made in the football game (about 200 gameplays).  Future predictions made at the U level focuses on the 4-5 important events to output rank future possibilities.          

Referring to FIG. 25, each predicted model has to consider predictions made by neighbors from its parent nodes and child nodes.  For example, G4 has to consider what G3 and G5 have predicted and the most important things in their lower levels (child nodes).  G4 also has to consider some of the predictions make in the upper levels such as P2 and P3 (parent nodes).  The parent nodes (P2, P3) contain a broader and general data about what are important objects/events contained between lower-level predicted models G2-G7.

FIG. 25

By doing predictions in a hierarchical structure, the important objects (large or microscopic) will be flushed out.  What if a blade of grass is responsible for the QB to trip and fall down.  The G level predicted models will show that the blade of grass is important in the football game.  In the P level predicted models, the QB and the blade of grass is important in the football game.  In the U level predicted model, the QB and the blade of grass was the turning point that made the Cowboys lose the entire game.

The hierarchically structured prediction tree, initially, generates an initial tree for the teams of virtual characters to work with.  This initial prediction tree was based on what prediction trees pre-existed in the prediction internet.  As the virtual characters work on predicted models, data in predicted models will change (data is added, deleted or modified).  The prediction work done by teams of virtual characters over a length of time will generate an optimal prediction tree.  The virtual characters’ predictions, flushes out the most important objects or events involved in the football game and structures and groups these important objects or events in a hierarchical manner (in other words, they modify the predicted models in the prediction tree).

This type of prediction method can work accurately with a football game.  The simplest outcome of a game is win or lose.  The predicted model U will look at all its lower levels and determine that P1, P2, P3, and P4 all agree that by the 3rd quarter, the Patriots is winning by a large margin compared to the Steelers.  The estimated score at the end of the 3rd quarter is 31-7.  The predicted model U can assume that the Patriots will win because it would take a miracle for the Steelers to make up 4 touchdowns in the 4th quarter.  Thus, in a general sense, the higher predicted models will have an accurate prediction of the future.  It’s up to the lower level predicted models to flush out rare events.   


<< content               next chapter >>




Home | Task Engine | Time Machine | Patents | Notes | Donation 

Copyright 2007 (All rights reserved)