Artificial intelligence and games: A long common history

By Bruno Rodríguez

The first appearance of artificial intelligence dates back to the 1950s, when John MCarthy
first used it, he was a computer scientist who ended up specializing in the subject. The first
paper on artificial intelligence is also about game theory, the one called "Programming a
computer for playing chess," a very impressive number to note is 10 ^ 123, which is the
number of possible chess games, if we start talking about who is the best players, are
usually machines, we have that for pacman it's a human, and for mrs pacman it's Maluuba,
an AI of microsoft.

So, what's Artificial Intelligence? We say, "Informatics discipline whose aim is to construct
smart programs or systems," if we adopt the concept of Turing to explain what it is to be
intelligent to a computer, we can say, "This is what makes it difficult to differentiate between
a task performed by a human being or a machine," there have been a lot of other people
trying to define what is artificial intelligence, and so on. At the same time, the scientist is
trying to simulate and emulate the behavior of the human brain, but we also need to
formalize logical thinking, and we have the first track of that with aristote, ​who said: “​All men
are mortal. Socrates is a man. Therefore, Socrates is mortal.” then in the 19 century the
mathematicians formalized the logic objects.

Nowadays, in the field of Artificial Intelligence, we want to define what the rational agent of
the convention that we have reached is "an entity that perceives and acts in its environment
to achieve its goals, to act well, where to act well is the action that maximizes its
performance, based on its abilities or beliefs (or knowledge)," and then, for every
environment or task, we try to find the best agent. Therefore, rationality depends on the
measurement of performance, the knowledge of the environment, the possible actions of the
agent and the sequences of perception of the agent, i.e. the order in which the information is
received.

Why it is so complex? Because usually the games/problems are NP-complete, and in the
learning process you need more than data, it is necessary to have knowledge (implicit or
explicit) to be able to learn, to reason, to justify a decision. Therefore it is needed to have
more complex structures to analyze and match differents structures with differents problems.

So how is usually done? The first step is modelling a real world situation, without AI one
should infer the values using various techniques, with AI the model will learn with the data
and optimize the model. The complexity of the model is highly variable, we can talk about
markov decision processes, bayesian networks, adversarial games, constraint satisfaction
problems, in some cases we have to model the situation as a model based on states and
graphs of states, that way, we can represent the world in each of the nodes of the graph
depending of some choices.
If we go deeper in the models based in states it is good to clarify the big three kinds we can
encounter:

1. Research problems: Everything is under control.

2. Markov pro+cess: There is randomness in the model.
3. Research against an adversary: There will be someone trying to sabotage the model.

We also have to note the models based on variables, the ones that the solutions to the
targeted problems consist in assigning values (the right ones) to variables and for this we
use the algorithms.

Returning to the historie of AI and the games, we can recall:

● Turochamp (1948): the first chess player created by Alan Turing

● Dartmouth conference (1956): the birth of AI
● Samuel’s Checkers AI (1956): IBM Checkers AI first demonstrated
● Bernstein’s Chess AI (1958): First fully functional chess AI developed
● Checkers AI Wins (1962): Samuel’s program wins game against person
● Mac Hack (1967): Chess AI beats person in tournament
● Zobrist’s AI (1968): First Go AI, beats human amateur
● Kaissa (1974): First world computer chess champion
● Backprop (1986): Multi-layer neural net approach widely know
● CNN (1989): Convolutional nets first demonstrated
● TD-Gammon (1992): RL and neural net based back-gammon AI shown
● Monte Carlo Go (1993): first research on Go with stochastic search
● Chinook (1994): Checkers AI draws with world champion
● NeuroGo (1996): ConvNet with RL for Go, 13 kyu
● Deep Blue (1997): IBM chess AI beats world champion
● MCTS Go (2006): French researchers advance Go AI with MCTS
● Crazy Stone (2008): MCTS Go AI beats 4 dan player
● Watson (2011): IBM watson wins at Jeopardy
● Zen19 (2012): MCTS based Go Ai reaches 5-dan rank
● DeepMind (2014): Google buys deep-RL AI company for $400Mil
● AlphaGo (2016): Deep Learning+MCST Go Ai beats top human
● AlphaStar - StarCraft2 (2019): The AI best the best players in SC 2

The main ideas are various, such as the Algorithm MINIMAX de John Vonn Neuman.Given
that in some cases it’s impossible to analyze all the possible games, we have to appeal to
some heuristics to calculate the solution or an approximation.

Why study the game theory?

Throughout our lives, we've had several games in which if we follow a certain strategy, we
increase the chances of winning, and we generally play to win, so it's a popular subject that
studies games because almost everyone wants to win, and not only that, the studies of
game theory and AI also can help us to develop other tools such as some works that
humans don't ​want to do, for example and adversary in a game o an assistance provider in a
website, we could keep adding examples. Some results from investigations in IA and game
theory, are very complex, but may have an obvious and accessible way to present the
findings so that they are closer to the general public.

In the categorization of games, we have the deterministic and non-deterministic games,

where the first ones we can calculate any solution depending on the first precedent state,
and in the second ones there exists randomness in the change from state to state, as well
we have games in the ones we don't know all the information in a current state, Poker is a
classic example of an imperfect information game because a player does not know whether
the opponent is holding a good or a bad pair of cards, and in others we do. Moreover we
need to consider the time complexity of the games, some of them ha​ve an ​exponential time
to calculate all the possible game results, such in chess, but in other cases is more complex
than that, like in Starcraft 2, there is more than 10^1685 posible states, for calculate the
complexity is a more complex process as in the algorithm analysis.

The next challenge in AI and game theory is to keep improving the tools that we already use
for content generation, AI adversaries for games, randomly generated scenarios in games,
movies, and continue developing new tools.