Mind - Learning


A virtual human needs to be dynamic, learning new things by experience and by more "traditional" methods such as reading. In Chapter 5 we present the following table to summarise how traditional learning definitions (given by Hopgood, 2011) might apply to a virtual human:

Rote

Converting an RSS news feed, or  text book page, into a set of RDF triples in the long term semantic memory, such as is done by the ConTEXT system (Khalili, 2014).

Advice

Using humans to do a task to provide a robot with a template to follow is a common approach in robotics (e.g. Rozo, 2013), often referred to as Learning from demonstration (LfD), and can be applied in virtual environments (McCormick, 2014).

Induction

Induction learning is similar to that found in most machine learning techniques, where the computer is trained on a set of examples (e.g. images or texts) and can then use these examples to identify a new image or piece of text (Kiesel, 2005).

Analogy

Perhaps one of the more challenging approaches for a virtual human since it must understand how two very different examples can actually be closely related.

Explanation

The virtual human tries to develop the explanations for why things happen. The PRODIGY system (Minton, 1989)  includes an implementation of an explanation-based learning system

Case-based

The virtual human acquires sets of cases that describe particular situations and particular solutions, and then when faced with a new problem finds the best matching solution Anthony (2017) describes how a case-based reasoning system was used with an intelligent agent to provide support for software programmers.

Explorative

Explorative learning is probably best exemplified by genetic learning algorithms. Here to find a solution the virtual human tries a variety of parameters at random, and then taking the best performing solution ‘breeds’ a new set, which it tries again, selects and breeds again and so on until it finds the best solution (Bredeche, 2012). Of course, there is no reason why a virtual human’s ‘exploration’ needs to be conducted in its ‘normal’ virtual world – it could readily be provided with a simulation of the simulation to try things out in first – so-called Projective Simulation (Mautner, 2014).



Having virtual humans that can learn could mean that there will be a significantly reduced need for programmers, and that instead a virtual human can be taught more by a subject matter expert, or even learn on its own. This creates a significant efficiency multiplier effect; a ‘generic’ virtual human with good learning capability could, quite rapidly, become a ‘specialist’ virtual human in one or more subjects/processes. As ‘unskilled’ virtual humans could be in almost limitless supply and could learn or be trained to undertake a wide variety of roles, from the trivial such as NPCs in video games, to the vital – such as caring for an aging population their impact on the need for human resource could be immense.