Reinforcement Learning

learn what to do from reward and punishment

optimize utility or fitness function with selected actions

learning about actions

"Reinforcement learning" in machine learning study

Multi-step decision-making in Markov decision process: given initial state, transition model T(s, a, s') and reward function R(s)
Policy: state-action mapping; optimal policy: maximum expected utility (sum of rewards)
Policy learning in a complete model: value iteration (calculating expected utility for each state) or policy iteration (evaluating and improving a working policy)
Passive learning: evaluate a fixed policy, and estimate transition model and reward function from reinforcement in trials
Active learning: exploitation vs. exploration tradeoff, bandit problem

Solomonoff: predicting future data given previous observations, provided the data is sampled from a computable probability distribution.
Hutter: maximizing reward in unknown environment with computable probability distribution.
To go through all Turing machines to find the most probable one, given observed data.
Using the algorithmic of a Turing machine to decide its prior probability.
Computability and computational complexity of the related models.

Evolutionary learning

various assumptions about the environment and the system

delayed feedback, credit/blame assignment

incomplete description, non-existing function, non-repeated decision

evolution and intelligence

solving problems with action sequences: static vs. dynamic [Computation and Intelligence in Problem Solving]

knowledge about action:

changing knowledge and resources restrictions may completely change the problem

intelligence and evolution: