Many artificial intelligence techniques focus on collecting unstructured material, such as free-form writing, photos, and audio, and extracting meaning from it in order to transform it into forms that are more valuable to humans. We haven’t done anything significant with the newly modified data, even though it is more helpful than the unstructured data it was produced from (the objects we have detected in photos, the intents and entities we have recognized in free text, the words we have identified in audio).
Fundamentally, optimization is about planning and problem resolution. It can discover the most effective way to do a task by using context provided by structured data from techniques such as Object Recognition or Natural Language Processing. This objective could be anything from improving one’s score in a computer or board game to closing a sale or streamlining a supply chain. Optimization employs a different strategy known as reinforcement learning, in contrast to previously covered approaches, which either depend on giving an algorithm pre-tagged data (supervised learning) or feeding it data and asking it to categorize that data (unsupervised learning).
Optimization appears to be repeated trial-and-error in its most basic form. The algorithm modifies its behavior slightly and checks to see if this moves it closer to its objective. If so, it keeps changing its behavior in the same ways. If not, it attempts an alternative. An optimization algorithm can learn to recall its previous actions and how they affected its objective using reinforcement learning. This enables it to begin creating strategies that result in greater results.
Let’s use an optimization algorithm that is learning to play Super Mario Bros. as an example. For those of you who haven’t seen a video game since 1983, the player controls Mario or Luigi as they move right (directionally) through stages by jumping from platform to platform. Certain foes can be defeated by jumping on them, or they can be avoided by hopping over them. You can’t just set an optimization algorithm’s goal to “beat Super Mario” and wait for it to succeed if you want to train it to play this game. This would take a very long period, and the algorithm would not have learned anything when it eventually won the game by accident. This is so that the algorithm is motivated to proceed in the right direction by providing it with short-term goals. The computer needs to learn how to move correctly and avoid obstacles, things that you and I would know instinctively. It is necessary to divide the challenging task of defeating Super Mario Bros. into a lot of smaller ones. It is necessary to give the game a goal that will eventually lead to defeat: maximizing score, rather than the ultimate objective of winning. Next, you configure a point system in which the algorithm gains points for each right movement, platform jump, and enemy kill. Object Recognition would enable this point system. The computer will learn techniques through hundreds or millions of trial and error to travel to the right, jump onto higher platforms, and eliminate adversaries. In the end, maximizing for these objectives will help you win the game.
However, what if we’re discussing an aggressive game—one in which players compete with one another—like DOTA 2 or Go? It is impossible for a human to teach an algorithm by having it play a million Go games against them. Firstly, it would be excessively lengthy (and extremely dull for the human participant). Secondly, the computer would only acquire skills that would make it slightly superior to the human opponent it faced. Generative Adversarial Networks (GANs) are useful in this situation. Thousands of times, the algorithm plays a slightly altered version of itself instead of competing against a human opponent. Following a game, both sides incorporate the lessons learned from the previous game into their own algorithms and launch a new one. Because GANs can play these games at a faster rate, they allow these algorithms to earn experience much faster. Because they aren’t comparing themselves to people, they also allow optimization algorithms to perform at levels that are higher than those of humans. This is how DeepMind trained its AlphaGo algorithm to defeat the world’s greatest human go player.
We’re done now! An optimization algorithm will eventually become skilled enough to defeat the greatest human players if you train it to maximize its score, right? Not quite, though. Short-term maximization is typically at the center of optimization approaches. This is particularly useful in scenarios where short-term gains will eventually translate into long-term success (such as Super Mario Bros.). In more intricate scenarios, such as those found in video games like DOTA 2 and StarCraft, making short-term sacrifices can pay off in the long run. This is partially the reason AI has struggled to outperform human player teams in games that demand the kind of strategic thinking that is difficult to model and take place in intricate, dynamic contexts.
To address this shortcoming, researchers have combined various systems in an effort to strike a balance between short- and long-term strategy. While one algorithm looks at the optimal move to make next, another considers the possible outcomes of the game as a whole. When combined, these two algorithms can determine the optimal route to success. A similar strategy was employed by the Libratus system, which defeated skilled poker players. One of its algorithms taught itself the game and then figured out what to do next by using reinforcement learning. A second concentrated on the game’s conclusion. A third algorithm would find trends in the past bets made by Libratus itself, and would also introduce unpredictability to confuse the other players.
What connection is there between card, board, and video games and how optimization may assist you manage your business? We bring up games while discussing optimization because academics evaluate the effectiveness of their algorithms on these kinds of games. Beating computer games is a short-term aim toward developing strong Optimization algorithms that can deal with the real world, much like we set the Optimization algorithm short-term goals to defeat Super Mario Bros. The ultimate goal is to train algorithms to function in complicated, uncertain, ever-changing situations rather than to create a system that can win a computer game.
Applications of optimization can be seen in real-world scenarios when a certain objective needs to be met. Facebook has already trained chatbots to haggle for basic goods just as well as humans can by using optimization techniques. Recommendations, shift rotation designs, and supply chain optimization are more applications of optimization. Optimization can be used to find the best course of action if a big decision can be divided into smaller judgments that can be optimized through trial and error.