Browse Prior Art Database

Plan Recognition with Negative Observations Disclosure Number: IPCOM000248476D
Publication Date: 2016-Dec-05
Document File: 3 page(s) / 33K

Publishing Venue

The Prior Art Database


Disclosed is a technique to use Artificial Intelligence (AI) in a solution to address negative observations in plan recognition systems.

This text was extracted from a PDF file.
This is the abbreviated version, containing approximately 40% of the total text.

Page 01 of 3

Plan Recognition with Negative Observations

Plan recognition techniques enable a system to infer the plans and goals of the agents
(i.e., executors of said plans and goals) based on observations. In plan recognition, the task is to find a likelihood or ranking of the plans and goals given a set of observations. These observations are often interpreted in the positive sense, that is, the observations indicate an occurrence of an event or presence of a signal that was detected by associated sensors. However, it is possible that that the technique has extra knowledge, such as not seeing a particular observation, that should be considered. This additional knowledge based on unseen factors is referred to as negative observations. A method is needed to address negative observations in plan recognition.

Related work considers the following three types of observations in addition to the regular observations:

• Noisy observations that are seen, but are not consistent or in context with the rest of the observations. The proposed solution considers discarding these observations.

• Missing observations that should have been seen, but are not due either to sensor malfunctioning or in the plan recognition problem. It is possible that the actions that make these observations true have not been executed yet. This is true in particular if the technique is predicting the future or trying to recognize the goal of the agent from the initial set of observations.

• Negative observations are not seen, implementers know the technique has failed to see this observation

Consider a simple kitchen example, in which the agent is making a meal, and observable actions include picking up a spoon or toasting bread. The task is to detect the agent's goal, and then in the cognitive assistant setting possibly intervene and help the agent achieve the goal. Now consider seeing that the agent is taking bread. At his point, it could be that the agent is going to make a turkey sandwich for lunch or dinner, or in the future take butter and toast the bread in order to make breakfast. The observations (e.g., butter for the breakfast scenario or turkey for the dinner or lunch scenario) are the missing observations because either it is an observation that was not seen, or it has not yet happened. At this point, all three scenarios are still possible. Now consider that, in addition, the observer knows that the toaster did not turn on; therefore, it is more likely that the agent is making a turkey sandwich.

In a more general sense, negative observations can be both over the past and the future. For example, if the observer knows that the agent does not like turkey or that the toaster is broken, then the observations associated with these are not seen. This additional knowledge comes from negative observations. The negative observations can then be used to improve the recognition of the goals or the plans of the agent,

without any other additions to domain knowledge.

The n...