AI planning involves teaching machines how to plan ahead. The process entails specifying problems informally using natural language, which is difficult for the planning system to understand. This is unlike the traditional approaches where the specification of the goal or problem to the planning system is assumed. Improving AI’s ability to understand problems through planning is an ambitious journey to make this technology part of everyone’s journey. For instance, proper planning will lead to the creation of systems that understand things such as our daily routines.
AI learning and planning
Human intelligence is defined by both learning and reasoning. While learning allows humans to take advantage of everyday activities, planning allows humans to plan future actions to achieve a specific goal. Therefore, planning complements learning and is one way of exploring the future, while learning involves getting knowledge from past experiences. Like humans, AI may encounter problems that are not related to past experiences and therefore require new solutions. Such problems are often difficult to solve this will require learning. Learning is used to make AI planning more efficient. This is just one way of ensuring planning and learning work together. Learning leads to the generation of macro-operators that increase the speed of the problem-solving process. For artificial intelligence, learning enables tuning of internal planning heuristics. This is critical in completing the definition of domain knowledge, which speeds up planning like hierarchical task networks.
Human vs. AI
Human intelligence is the point of reference when comparing the capabilities and the potential of AI. Although human intelligence is not necessarily the ground truth and AI does not entirely mirror human intelligence, the comparison allows us to get valuable insights about the parts of AI that are more developed and those that need more attention and improvement in the future. In the sub-area of learning, for example, the view out there is that AI is fundamentally different from human learning.
Unlike humans who learn from very few examples, AI requires many examples to learn. Humans have good generalization capabilities from just a few examples. On the contrary, AI requires many examples, which can be attained by availing large data portions. AI is said to be greedy, shallow, and brittle because of the vast amounts of data it requires to learn. It is shallow because neural networks have a narrow knowledge of commonsense and brittle because it has a limited generalization power. Understanding these differences between humans and machines helps define the objectives of AI and what it needs to achieve.
Unlike machines, humans have an excellent way of formulating planning problems. Humans know how to choose the details and the right abstraction levels around us and what to ignore at a specific level of abstraction. This is something that is not always easy for machines. Therefore, humans have an advantage in recognizing a problem and expressing it.
Conclusively, planning does not always require training or even the availability of vast amounts of data. Instead, it requires a combination of planning and learning. However, the goal and the size of the input can change the problem instance to another. The bottom line is understanding what and how important a problem is. That is intelligence.