Problem Solving and Goal-Driven Learning
There is a growing body of research into how a learner’s (human or machine) goals can greatly influence the learning process. Study has taken place in fields such as cognitive science, psychology, education and, of most interest to us, artificial intelligence. Prior to the interest in goal-driven learning, most studies in this field focused on providing estimated functions based on limited inputs and outputs, without concern for the learning goal.
The underlying principal of the goal-based methodology is that learning is, for the most part, a strategic and active attempt on the part of the learner to identify and solve a particular problem or set of problems within the context of the tasks, goals, prior knowledge or expertise available locally as well as opportunities provided by the problem-space for learning. The learning process should be guided by good decision making in regards to what information is needed to achieve a particular goal-state because the value of what is learned is dependent on what impact that learning has on achieving a goal.
In a person, a goal can be directly linked to their ambition or determination to achieve some end. A goal is a desired outcome. A machine can seem determined because it will persevere with a task until it is finished but this is an illusion and a machine can’t desire anything. Another aspect of a goal is that it must be a tangible outcome; you must be able to describe the desired finishing state or else how would you know if you had achieved it?
We can then say that for a machine to have a goal it must have two fundamental things; a description of what the end state should be and the ability to persist until it reaches this state or until it has exhaustively concluded that reaching the end state is impossible. A goal driven system treats all elements it encounters as objects to exploit, ignore or avoid in order to achieve an outcome that does not yet exist. When a goal driven system is presented with a situation which differs from the end-state it attempts to reduce the differences between the actual and the desired state recursively until no differences exist.
When we undertake to solve a problem we imagine creative solutions based on either expertise or common sense but it is hard for us, as humans, to conceptualize the idea of creativity in a machine. A computer carries out only the instructions that it is programmed with and therefore cannot innovate. However a computer can produce unexpected solutions not envisioned by the original programmer by using a trial and error method where the only required input is that it can recognize when the problem is solved. This is also known as the Generate and Test method.
If we imagine constructing a house from a set number of building blocks with different shapes for the walls and roof. In the first phase we simply generate any arrangement of all the blocks, this would be simple...