Hello I am Nitsan Soffair, A Deep RL researcher at BGU.

In my Deep reinforcement-learning course you will learn the newest state-of-the-art Deep reinforcement-learning knowledge.

You will do the following

1. Get state-of-the-art knowledge regarding

   1. Model types

   2. Algorithms and approaches

   3. Function approximation

   4. Deep reinforcement-learning

   5. Deep Multi-agent Reinforcement-learning

2. Validate your knowledge by answering short and very short quizzes of each lecture.

3. Be able to complete the course by ~2 hours.

Syllabus

1. Model types

   1. Markov decision process (MDP)

      A discrete-time stochastic control process.

   2. Partially observable Markov decision process (POMDP)

      A generalization of MDP in which an agent cannot observe the state.

   3. Decentralized Partially observable Markov decision process (Dec-POMDP)

      A generalization of POMDP to consider multiple decentralized agents.

2. Algorithms and approaches

   1. Bellman equations

      A condition for optimality of optimization of dynamic programming.

   2. Model-free

      A model-free algorithm is an algorithm which does not use the policy of the MDP.

   3. Off-policy

      An off-policy algorithm is an algorithm that use policy 1 for learning and policy 2 for acting in the environment.

   4. Exploration-exploitation

      A trade-off in Reinforcement-learning between exploring new policies to use existing policies.

   5. Value-iteration

      An iterative algorithm applying bellman optimality backup.

   6. SARSA

      An algorithm for learning a Markov decision process policy

   7. Q-learning

      A model-free reinforcement learning algorithm to learn the value of an action in a particular state.

3. Function approximation

   1. Function approximators

      The problem asks us to select a function among a well-defined class that closely matches ("approximates") a target function in a task-specific way.

   2. Policy-gradient

      Value-based, Policy-based, Actor-critic, policy-gradient, and softmax policy

   3. REINFORCE

      A policy-gradient algorithm.

4. Deep reinforcement-learning

   1. Deep Q-Network (DQN)

      A deep reinforcement-learning algorithm using experience reply and fixed Q-targets.

   2. Deep Recurrent Q-Learning (DRQN)

      Deep reinforcement-learning algorithm for POMDP extends DQN and uses LSTM.

   3. Optimistic Exploration with Pessimistic Initialization (OPIQ)

      A deep reinforcement-learning for MDP based on DQN.

   4. Value Decomposition Networks (VDN)

      A multi-agent deep reinforcement-learning algorithm for Dec-POMDP.

   5. QMIX

      A multi-agent deep reinforcement-learning algorithm for Dec-POMDP.

   6. QTRAN

      A multi-agent deep reinforcement-learning algorithm for Dec-POMDP.

   7. Weighted QMIX

      A deep multi-agent reinforcement-learning for Dec-POMDP.

Resources

• Wikipedia

• David Silver's Reinforcement-learning course