Written by: Lee Wee Sun, Oct 2019
Finding the optimal solution for most interesting AI problems is computationally intractable. Herbert Simon coined the term satisficing for how decision makers make decisions when they are unable to find the optimal solution. So, what should AI researchers do?
One reason AI problems are often considered intractable is because we try to solve the worst case problems. Perhaps we should solve the typical cases and try to do well on average. Machine learning provides excellent tools for solving for the average case. Machine learning algorithms typically try to work well on a training set, and if the learned solution generalizes well, it would work well for the true distribution.