Charu C. Aggarwal > Quotes

“Another problem with building knowledge bases is that they can work only in highly interpretable settings. This property is embedded into the inherent nature of knowledge bases, as humans construct them with their semantic insights. While an interpretable knowledge base might seem to be a virtue at first glance, it is also a problem because many intelligent decisions made by humans cannot be easily enunciated in words. For example, a chess grandmaster or an expert player of the game of Go might sometimes be unable to concretely explain why they choose a particular move, beyond the fact that their experience from previous games translates into an intuitive but hard-to-enunciate understanding of favorable spatial patterns. Trying to handcraft this type of intuitive knowledge into a semantically interpretable board evaluation function is often a source of inaccuracy, as it misses the intangibles in the decision-making process.”
― Charu C. Aggarwal, Artificial Intelligence: A Textbook

“The Monte Carlo tree search method is naturally suited to non-deterministic settings such as card games or backgammon. Minimax trees are not well suited to non-deterministic settings because of the inability to predict the opponent’s moves while building the tree. On the other hand, Monte Carlo tree search is naturally suited to handling such settings, since the desirability of moves is always evaluated in an expected sense. The randomness in the game can be naturally combined with the randomness in move sampling in order to learn the expected outcomes from each choice of move.”
― Charu C. Aggarwal, Artificial Intelligence: A Textbook

“The primary strength of inductive learning systems lies in their being able to capture the inexpressible part of human cognition. For example, when a child learns a language through speaking, she does not start by memorizing the rules of grammar, but by “picking up” the language in a way that cannot be fully explained even today. Therefore, the child grows in linguistic ability by learning through examples. The child might occasionally receive some knowledge from her parents, such as specific concepts in grammar or vocabulary, but it is rarely the primary form of learning for native languages.”
― Charu C. Aggarwal, Artificial Intelligence: A Textbook

“increased computational power, which almost always improves inductive learning methods far more than deductive reasoning methods. Indeed, many of the exciting results in artificial intelligence in the previous decade have started to show that the classical school of thought in artificial intelligence (i.e., deductive reasoning) had serious limitations that were often overlooked in the early years (at the expense of inductive learning methods).”
― Charu C. Aggarwal, Artificial Intelligence: A Textbook

“The ability to go beyond human domain knowledge is usually achieved by inductive learning methods that are unfettered from the imperfections in the domain knowledge of deductive methods.”
― Charu C. Aggarwal, Artificial Intelligence: A Textbook

“A key point is that an increased number of attributes relative to training points provides additional degrees of freedom to the optimization problem, as a result of which irrelevant solutions become more likely. Therefore, a natural solution is to add a penalty for using additional features. Specifically we can add a penalty for each parameter w i, which is non-zero. One can express this penalty using the L 0-norm of the vector”
― Charu C. Aggarwal, Artificial Intelligence: A Textbook

“Monte Carlo trees tend to explore a few promising branches deeper based on evaluations from previous experience, whereas minimax trees explore all unpruned branches in a roughly similar way. The human approach to chess is similar to the former, wherein humans evaluate a small number of promising directions of play rather than exhaustively considering all possibilities. The result is that the style of chess from Monte Carlo tree search is more similar to humans than that from minimax trees. The programs resulting from Monte Carlo trees can often take more risks in game playing, if past experience has shown that such risks are warranted over the longer term. On the other hand, minimax trees tend to discourage any risks beyond the horizon of tree exploration, especially since the evaluations at leaf levels are imperfect.”
― Charu C. Aggarwal, Artificial Intelligence: A Textbook

“Monte Carlo tree search makes predictions by generating examples (rollouts) and then makes a prediction based on the experience of this algorithm with these examples. Modern variations of Monte Carlo tree search further combine tree search with reinforcement learning and deep learning in order to improve predictions. Deep learning is used in order to learn the evaluation function instead of using domain knowledge. For example, AlphaZero uses Monte Carlo tree search for rollouts, while also using deep learning to evaluate the quality of the moves and guide the search. The Monte Carlo tree together with the deep learning models form the hypothesis used by the program in order to move moves. Note that these hypotheses are created using the statistical patterns of empirical behavior in games.”
― Charu C. Aggarwal, Artificial Intelligence: A Textbook