Reinforcement learning (RL): sequence of decisions
Nature: the actual situation changing methods in order to achieve the optimal solution it
The reason:
1) incremental update parameters (Incremental Parameter Adjustment), from input to output, there is a gradient descent is completed (in increments as small, slow learning process (Each increment will need to be very small, they are not to let the new information learned , the lessons learned previously covered (this is called "catastrophic interference")))
DRL Solution: Add RNN am thinking, information learned from the events of the past years, can come in handy once
2) weak inductive bias (Weak Inductive Bias), any learning process, should face a "bias - variance trade-off."
A given amount (bias), AI junction consequences quickly get
DRL solution: learn from past experience, to a narrow range
Evolutionary Algorithms:
The value of fitness function, is often the end result we want, in fact, just to go to pick iteration point in the function space, to search for the point that approximate the optimal value
Differences: Although reinforcement learning iteration is also sampled in order to function in space, but it is a cumulative approach to the optimal value. So we are concerned about the cumulative
Bayesian optimization:
Prior knowledge approximate the unknown target function of the posterior distribution, concerns over various parameters