In the past 2017 years, deep learning technology to flourish, AlphaZero begin from "zero" rapid development in a variety of chess competition, DeepStack and Libratus Texas poker win over human masters, GAN derived variety, speech synthesis from laboratory to product, Vicarious probability model and put forward a new human CAPTCHA code. These exciting progress will be smart technology from the laboratory to industry and application level, "artificial intelligence" and "deep learning" concept has also entered the public sight and become a popular word.
It's easy to confuse artificial intelligence with machine learning (and even deep learning), and feel that they are the same. This is a big misunderstanding. In addition to popular neural network model, the history of Artificial Intelligence can be traced back to 1956, during the Dartmouth called "Dartmouth Summer Research Project of Artificial Intelligence (Dartmouth Summer Research Project on Artificial Intelligence)" seminar, "Artificial Intelligence (Artficial Intelligence)" the term (AI) was put forward for the first time, also on the form of Artificial Intelligence Research field.
The "ai" discussed in this report mainly refers to the intelligent (also known as Machine Intelligence) that can be embodied by machines. In the field of academic research, it is the intelligent agent that can perceive the surroundings and take action to achieve the best possible result.
In general, the long-term goal of AI is to implement general artificial intelligence (AGI), which is seen as "strong AI". AGI performs far more than the average machine when dealing with cross-domain issues, and it can handle multiple tasks at once. Weak AI, also known as "narrow AI", cannot solve problems that have not been seen before, and its capabilities are limited to specific areas. But the exact definition of AGI is still unclear to artificial intelligence experts and scientists. A common way to distinguish between strong ai and weak ai is to conduct professional tests, such as coffee tests, Turing tests, robot college students and job tests.
The "technology" discussed in this report is a broad concept that includes methods, algorithms and models used in the field of artificial intelligence, and we will use the term "technology" to refer to these three.
Work related to artificial intelligence can be traced back to the 1940 s, when Warren McCulloch and Walter Pitts research shows that can be executed by interconnection of the neural network computation, Donald Hebb also demonstrates Hebb learning (Hebbian learning). In 1956, when the term "artificial intelligence (AI)" was formally established, corresponding research began to develop rapidly. At that time, artificial intelligence was mainly used for solving problems. Although during the 70 s and 1960 s, the development of the neural network is slow speed and limited progress has received criticism, but the presence of the expert system to maintain the interest in artificial intelligence and the growth of the related research. Soon after, in the 1980 s, because most of the research of artificial intelligence are unable to implement their original commitment to funding for artificial intelligence research is also going to other areas, the academic research of artificial intelligence into the so-called "Winter Winter (AI) artificial intelligence. Luckily, the back propagation neural network to link activists to reinvent back to the stage. In the 1990 s, researchers began to apply probability model more scientific methods, such as; at the same time, the support vector machine (SVM) in many areas all over and replaced the neural network. Soon into the new 21st century, the era of big data, which helps researchers have developed a variety of learning algorithm, makes deep learning is booming in recent years. More and more attention to artificial intelligence research team, focused on artificial intelligence of top class meeting (such as AAAI, IJCAI, AUAI and ECAI) on paper is also growing rapidly.
The history of search algorithms began during the incubation period of artificial intelligence in the 1950s, when the focus was on solving problems. The term "search" in artificial intelligence mainly refers to the search algorithms that drive computer/smart bodies, which enable computers to solve problems in a human way. Search algorithms can generally be divided into two categories: uninformed search and information (in-formed) search. The information heuristic search is quite popular because it can quickly find solutions based on some instructions. This category contains one of the most popular search algorithms: A* search. A* search can be defined as A (greedy) best first search. Other heuristic (information) Search algorithms also include recursive optimal Search (RBFS), Simplified limited Memory A*(Simplified memory-bounded algorithm/SMA *), monte carlo tree Search (MCTS), and Beam Search. No information search algorithm includes width priority search, depth priority search, depth limit search, two-way search and iterative deepening search. The heuristic search algorithm has a local search algorithm, which is especially effective for constraint satisfaction (CSP) [26]. Popular local search algorithms include mountain climbing algorithms, simulated annealing algorithms, local cluster search and genetic algorithms (variations of random cluster search).
For the constraint satisfaction (CSP) mentioned above, the solution is expressed by a variable that satisfies a set of specific constraints. This is a very hot research topic in the field of artificial intelligence and operational research. Local search algorithms can be used to solve such problems, and local search of minimum conflict heuristic has been a great success in solving CSP. In addition, related technologies include backtracking algorithms (a deep priority search algorithm) and tree decomposition techniques.
In addition to the traditional single agent search optimization problem, and focus on "for different party planning" search algorithm (such as a competitive game), the algorithm is called against search algorithm (adversarial search algorithm). Classic algorithms to combat search problems include Minimax (Minimax) search and alpha-beta pruning.
Current stage: because the anti-search technology has been widely used in game gaming for a long time, the development phase of the search algorithm for problem solving is determined as the application phase.
Bottleneck: although different search algorithms have applications, there is no clear direct commercial solution to search algorithm technology.
Future: with the development of deep learning, integration with search algorithms is promising (for example, MCTS using intensive learning). More recently, it has been a big hit with the creation of an anti-search network (GAN). Predictably, there will be more smart bodies in many different applications.