Generative event sequence simulator with probability estimation

1. A non-transitory machine-readable medium storing instructions that when executed by one or more processors effectuate operations comprising:

obtaining, with a computing system, a generative artificial intelligence (AI) model, the generative AI model having been trained to output a predicted sequence of events;

inputting, by the computer system, a first sequence of at least one event to the generative AI model, the at least one event being associated with a positional encoding representing a physical location in a region of physical space at which the at least one event occurred;

predicting, with the generative AI model, a second sequence of at least one event subsequent to the first sequence of at least one event based on the first sequence of at least one event and the positional encoding; and

storing, with the computer system, the second sequence of at least one event in memory.

2. The medium of claim 1 , wherein the positional encoding represents at least one spatial dimension of the physical location in the region of physical space with a vector having a plurality of scalars corresponding to a plurality of different frequencies of a periodic wave function.

3. The medium of claim 2 , wherein the generative AI model is further trained to generate predicted physical locations in the region of physical space for at least one of an event, an actor, or an event object of the second sequence.

4. The medium of claim 1 , wherein the positional encoding is a sinusoidal positional encoding; wherein the positional encoding is a concatenation of positional encoding corresponding to multiple dimensions of the physical location; wherein the first sequence of at least one events is comprised of at least one event from a set of tracked events; and wherein the region of physical space is bounded and predefined.

5. The medium of claim 1 , wherein inputting the first sequence of at least one event comprises inputting multiple positional encodings for at least one event of the first sequence, the multiple positional encodings generated based on physical locations of multiple of one or more events, one or more actors, or an event object.

6. The medium of claim 1 , wherein:

the generative AI model is a transformer with multi-headed attention.

7. The medium of claim 1 , wherein some of the events of the first sequence do not include positional encodings representing physical locations in the region of physical space.

8. The medium of claim 1 , wherein the generative AI model is trained to predict events in a sporting match, weather events, crop yield events, crowd behavior events, forest fire events, crime events, material deformation or failure events, corrosion or oxidation events on metal surfaces, and maintenance events in industrial process equipment.

9. The medium of claim 1 , wherein the generative AI model is a transformer trained to predict events in a match that is a contest between two or more teams, wherein each team comprises one or more actors, wherein predicted events correspond to a set of tracked events identified as possible occurrences in the match, and wherein at least some of the events comprise events corresponding to one or more actors.

10. The medium of claim 9 , wherein the first sequence of at least one event comprises events which have already occurred in the match and wherein the second sequence of at least one even comprises a predicted sequence of events in the match, wherein the inputs also include descriptions of actor injuries.

11. The medium of claim 1 , wherein:

the generative AI model comprises a generative transformer;

inputting the first sequence to the generative transformer further comprises inputting a prompt to the generative transformer; and

the generative transformer is further trained to generate the predicted sequence of events based on the prompt.

12. The medium of claim 11 , wherein the prompt comprises information about an actor causing at least one event, wherein the information is available before the at least one event caused by the actor causes the event.

13. The medium of claim 11 , wherein the prompt and/or the first sequence of at least one event are tokenized as input embeddings and wherein the input embeddings are associated with a plurality of positional encodings corresponding to physical locations in the region of physical space.

14. The medium of claim 1 , wherein the generative AI model is a non-deterministic model, and wherein predicting comprises predicting multiple sequences and determining population statistics based on the multiple sequences to estimate a likelihood of a specified event or class of events.

15. The medium of claim 1 , further comprising steps for setting odds on an event occurring in a sporting match.

16. The medium of claim 1 , further comprising steps for encoding physical location at which events occur in the region of physical space.

17. The medium of claim 1 , further comprising steps for training the generative AI model.

18. The medium of claim 1 , wherein the first sequence events are obtained by processing a video feed of at least part of the region of physical space with a computer vision model trained to detect at least some of the first sequence of events.

19. The medium of claim 1 , wherein:

the generative AI model is nondeterministic;

the generative AI model is configured to determine, based on the prediction, joint probabilities of a set of a plurality of events occurring.

20. The medium of claim 1 , the operations comprising:

detecting, based on the prediction, that an anomalous event has occurred; and

causing an alert to be presented characterizing the anomalous event.

21. The medium of claim 1 , wherein the at least one event is associated with an additional positional encoding representing a position of the event in the first sequence and predicting the second sequence comprises predicting the second sequence of at least one event subsequent to the first sequence of at least one event based on the first sequence of at least one event, the positional encoding, and the additional positional encoding.

22. A processor-mediated method comprising:

obtaining, with a computing system, a generative artificial intelligence (AI) model, the generative AI model having been trained to output a predicted sequence of events in response to receiving predicate events, wherein the predicted sequence of events are not natural language text tokens, and the events are part of an at least partially stochastic process that occurs over a region of physical space;

inputting, by the computer system, a first sequence of at least one event to the generative AI model, the at least one event being associated with a positional encoding representing a physical location in the region of physical space at which the at least one event occurred;

predicting, with the generative AI model, a second sequence of at least one event subsequent to the first sequence of events based on the first sequence of at least one event and the positional encoding; and

storing, with the computer system, the second sequence of at least one event in memory.

23. The method of claim 22 , wherein the positional encoding represents at least one spatial dimension of the physical location in the region of physical space with a vector having a plurality of scalars corresponding to a plurality of different frequencies of a periodic wave function.

24. The method of claim 23 , wherein the generative AI model is further trained to generate predicted physical locations in the region of physical space for at least one of an event, an actor, or an event object of the second sequence.

25. The method of claim 22 , wherein the positional encoding is a sinusoidal positional encoding; wherein the positional encoding is a concatenation of positional encoding corresponding to multiple dimensions of the physical location; wherein the first sequence of at least one events is comprised of at least one event from a set of tracked events; and wherein the region of physical space is bounded and predefined.

26. The method of claim 22 , wherein inputting the first sequence of at least one event comprises inputting multiple positional encodings for at least one event of the first sequence, the multiple positional encodings generated based on physical locations of multiple of one or more events, one or more actors, or an event object.

27. The method of claim 22 , wherein:

the generative AI model is a transformer with multi-headed attention.

28. The method of claim 22 , wherein some of the events of the first sequence do not include positional encodings representing physical locations in the region of physical space.

29. The method of claim 22 , wherein the generative AI model is trained to predict events in a sporting match, weather events, crop yield events, crowd behavior events, forest fire events, crime events, material deformation or failure events, corrosion or oxidation events on metal surfaces, and maintenance events in industrial process equipment.

30. The method of claim 22 , wherein the generative AI model is a transformer trained to predict events in a match that is a contest between two or more teams, wherein each team comprises one or more actors, wherein predicted events correspond to a set of tracked events identified as possible occurrences in the match, and wherein at least some of the events comprise events corresponding to one or more actors.

31. The method of claim 30 , wherein the first sequence of at least one event comprises events which have already occurred in the match and wherein the second sequence of at least one even comprises a predicted sequence of events in the match, wherein the inputs also include descriptions of actor injuries.

32. The method of claim 22 , wherein:

the generative AI model comprises a generative transformer;

inputting the first sequence to the generative transformer further comprises inputting a prompt to the generative transformer; and

the generative transformer is further trained to generate the predicted sequence of events based on the prompt.

33. The method of claim 32 , wherein the prompt comprises information about an actor causing at least one event, wherein the information is available before the at least one event caused by the actor causes the event.

34. The method of claim 32 , wherein the prompt and/or the first sequence of at least one event are tokenized as input embeddings and wherein the input embeddings are associated with a plurality of positional encodings corresponding to physical locations in the region of physical space.

35. The method of claim 22 , wherein the generative AI model is a non- deterministic model, and wherein predicting comprises predicting multiple sequences and determining population statistics based on the multiple sequences to estimate a likelihood of a specified event or class of events.

36. The method of claim 22 , further comprising steps for setting odds on an event occurring in a sporting match.

37. The method of claim 22 , further comprising steps for encoding physical location at which events occur in the region of physical space.

38. The method of claim 22 , further comprising steps for training the generative AI model.

39. The method of claim 22 , wherein the first sequence events are obtained by processing a video feed of at least part of the region of physical space with a computer vision model trained to detect at least some of the first sequence of events.

40. The method of claim 22 , wherein:

the generative AI model is nondeterministic;

the generative AI model is configured to determine, based on the prediction, joint probabilities of a set of a plurality of events occurring.

41. The method of claim 22 , further comprising:

detecting, based on the prediction, that an anomalous event has occurred; and

causing an alert to be presented characterizing the anomalous event.

42. The method of claim 22 , wherein the at least one event is associated with an additional positional encoding representing a position of the event in the first sequence and predicting the second sequence comprises predicting the second sequence of at least one event subsequent to the first sequence of at least one event based on the first sequence of at least one event, the positional encoding, and the additional positional encoding.