Department of Electrical Engineering
Control Robotics and Machine Learning Lab
Technion - Israel Institute of Technology
המעבדה לבקרה רובוטיקה ולמידה חישובית
Goal - Solving sequential decision-making problems in high dimensional enviroments.
Sequential decision making -
An agent is interacting with an environment by performing actions (a) and observing rewards (r) and states (s). The agent's goal is to choose actions in order to maximize its total future reward.
Approximations -
Real world problems suffer from the exponential growth in the size of the state space - the well known "curse of dimensionality". Therefore, in order to solve a real-world problem we need to use approximations from the 'realm' of supervised and deep learning.
We focus on two domains:
Minecraft
Minecraft is a role-playing game where the goal of a player is to survive and prosper. The player can build anything from homes to farms. This game provides a challenging framework for Reinforcement Learning problems, and more generally artificial intelligence, due to its high dimensionality and stochasticity.
The arcade Learning Enviroment (ALE) -
An interface to hundreds of Atari 2600 game environments, each one different, interesting, and designed to be a challenge for human players. ALE presents significant research challenges for reinforcement learning, model learning, model-based planning, imitation learning, transfer learning, and intrinsic motivation
Machine Learning branches
Supervised Learning
Unsupervised Learning
The computer is presented with example inputs and their desired outputs.
The goal is to learn a general rule which maps inputs to outputs.
Examples: Classification, Regression.
No labels are given to the learning algorithm.
Finds hidden patterns in its input.
Examples: Clustering, Anommaly detection, Dimensionality reduction.
Reinforcement Learning (RL)
Deep Learning (DL)
Select actions to maximise total future reward.
There is no supervisor, only a reward signal.
Actions may have long term consequences, reward may be delayed.
Examples: Teaching a Robot to walk, controlling a power station, learning to play Backgammon or Atari.
Branch of machine learning that aims to mimic the brain.
Multiple layers of linear and nonlinear processing units in one deep joint architecture.
Efficient GPU implementations.
Related to all types of learning.