Intensive AI is based on the use of combinatorial agents that learn on limited amounts of data by recognizing patterns and combining them for making decisions on how to solve problems and to learn further.

The learning speed is determined by the ability of the agent to build hierarchies of nested decision-making patterns and to go away from these patterns when the character of the problems changes. Here the go-away is based not on forgetting but on finding features in the problem definitions that can be used for selecting right patterns, which also results in a pattern. Finding patterns in data generated in the process of building and transforming other patterns is known as a concept of meta-learning.

So the main functions of combinatorial agent in the intensive AI approach are:

• Learning as finding patterns in differently structured data,
• Generalization as effective application of the patterns to new data not used in training,
• Meta-learning as finding patterns in data generated in its internal process of building and transforming patterns.

There are only three ways to accelerate solving a training problem:

• To use a priori information about the training problem to build the learning agent,
• To provide the agent with ready examples of solutions during the training,
• To supervise the training process

The latter two ways correspond to training, while first one relates to the internal design of the learning agent. It’s the prior information used for acceleration of learning that usually narrows down the application of the trained agent.

The learning algorithms of combinatorial agents use as little prior information about the structure of data the agent will encounter with as possible. The only prior information the algorithms use is that there are patterns in decision making, that the patterns can be reused and nested in some ways to form hierarchies. All the patterns and their functions are derived from the training data on the basis of these prior assumptions.

These assumptions are enough for the training acceleration, while on the other hand they do not limit practical applications because there are patterns and hierarchies in any scope of human activity and the nature itself proves to be structured in many aspects.

Machine understanding is also an effect that emerges in communications with an agent trained to relate the use of the patterns to tokens in some sign system, where the relation itself is patterned as well.

There are two aspects of the power of mind, the one is computational and the other is algorithmic. Parallelizing tasks and scaling up data storage are little problems nowadays. The main question is the algorithm that results in learning to understand and abstract thinking. Combinatorial AI is a part of the intensive algorithmic approach elucidating the routines of intelligence in contrast to scaling and hardware acceleration, which are the methods of the extensive way.

So the algorithm of the combinatorial agent is purposefully made on the basis of CPU to get away from the restrictions imposed by GPU computation method and it runs effectively when only some of its procedures are executed in parallel.

The advantage of the computation of the the learning algorithm of the combinatorial agent on CPU compared to GPU-based neural networks grows as the training problems become more and more complex.

The algorithm of the combinatorial agent allows forming AI with new abilities, which are critical in many applications:

• Learning on limited amounts of data, which is usually the case when the learning shifts from training examples to solving real practical problems,
• Machine understanding as the property of communication between developer or user with AI, which accelerates further learning, brings about precise problem statements and requirement specifications, find and share new new .
• Abstract thinking as the ability to solve problems of finding methods applicable in the domain the agent was trained on.

The useful functions of combinatorial AI result not only from the learning algorithms but also from special training programs made in accordance with the hierarchical train-and-use methodology, which comprises

• Routines regulating the choice of the problems and the way of communication with the agent for speeding up training process
• Indicators to monitor how agent’s learning mechanisms and training methods correspond to the goal of machine understanding and abstract problem solving

The training runs from simple to complex problems, from the subject-specific basics to the higher abstract levels.