Integrated Learning (IL)

Program Manager: Dr. Todd Hughes

Background:

The motivating challenge problem for Integrated Learning is to learn (symbolic) hierarchical task models, complex processes, or generalized plans by watching a human user perform a task just once. Figure 1 shows an example taken from a physical domain in which the learner must produce a generalized task model for assembling an object after being show the assembly process, once, by a human user.

Learn A Complex Plan From Single Set Of Observations

Click Here for Figure One

Once such a model is learned, it can be analyzed or incrementally extended by subsequent learning or by additional planning/reasoning activities. With such models systems can automate task performance or provide intelligent, contextually specific instruction to a human user. For instance, a person assembling an object might wind up in a dead-end state that does not lead to a goal. Using such a structure, an Integrated Learning system might back the human up to a prior state and then direct him/her down a different path to goal achievement.

The expression learning as an integrated problem solving process identifies two important ideas: First, that learners in this program will be integrated in a meaningful two-way fashion with other components in a cognitive system and able to utilize their knowledge and their reasoning in the learning process. Secondly, learners will regard learning as a problem to solve rather than a rote series of steps or operations to perform. On the latter point, this means that Integrated Learners will:

1. have explicit learning goals and formulate plans to achieve them;
2. keep track of what they don’t know and what they need to know;
3. form hypothesis and track uncertainties associated with them;
4. be both opportunistic and process driven in their control; and
5. assemble knowledge from multiple sources and build on that knowledge.

This is not a complete or exhaustive list. From an executive level, Integrated Learners attempt to “figure things out” rather than execute a set of predefined/static algorithmic steps. This approach will yield learning systems that are more flexible; where the learners are able to use many different sources of information, process information in many different forms, and proactively work with reasoning and simulation components to generate desired information. Integrated Learners will also be more robust – tolerant of errors in information and tolerant against missing information because the learner can draw on whatever information or reasoning is available to support learning and can use multiple sources to corroborate or negate hypotheses.

An example Integrated Learner is shown in Figure 2. Note that the learner incorporates world knowledge, domain knowledge, several types of sophisticated reasoning and simulation, and a module for conventional statistical machine learning algorithms. These other components or modules are tools that the Integrated Learner employs during learning to generate knowledge that it needs to achieve its learning goals. This interconnected view is very different from the algorithmic focus of statistical machine learning algorithms where the algorithm has one "input pipe" through which training data (of a very specific form) flows. Integrated Learners have many “pipes” and must be able to manipulate many different forms of information and even trade off different types of information and reasoning. Integrated Learners can also interact directly with a human user to fill learning information needs. However, the learner must perform a cost/benefit trade-off analysis before invoking the human as the human interaction may be a more “expensive” option than other computational options the learner may have.

An Example Integrated Learner for Physical Domains

Click Here for Figure Two