My solution of Practical Reinforcement Learning by National Research University Higher School of Economics via Coursera.

The GitHub version of the course.

• Reinforcement Learning
  • Multi-armed bandit
  • Decision process & applications
• Black box optimization
  • Markov decision process (MDP)
  • Crossentropy method
    • rollout experiences with stochastic policy $\pi_i$
    • store <s,a> pair of better reward as elite
    • update policy $\pi_{i+1}$ proportional to the occurrences of <s,a> pair in elite group
  • Approximate crossentropy method *
• Evolution strategies

• Practice basic interface of OpenAI gym environment.

• Solved 'Taxi-v3' with crossentropy method with a table representation of policy.

• Approximate the policy of a continuous state space game 'CartPole-v0' by multi-layer neural network (MLP)
• Train the MLP with elite <s,a> pair selected by crossentropy method

• Reward
  • Reward design
  • Discount reward
• Bellman equations
  • Value functions
  • State action value function
• Generalized policy iteration
  • Policy evaluation & improvement
  • Policy and value iteration

• Implemented value iteration algorithm to solve MDP problem
  • $V_{i+1}(s) = max_a \Sigma_{s'} P(s'|s,a)\cdot [r(s,a,s')+\gamma V_i(s')]$

• Model-free learning
  • Monte-Carlo & temporal difference; Q-learning
  • Exploration vs exploitation
• On-policy vs off-policy

• implement vanilla Q-learning algorithm
  • $Q(s,a) \leftarrow (1-\alpha)Q(s,a) + alpha *(r+\gamma V(s'))$
• on both discrete state space environment 'Taxi-v3' and discretized continuous state space environment 'CartPole-v0'

• implement Expected Value SARSA algorithm
  • sample <s,a,r,s'> from environment
  • $\hat{Q}(s,a) = r(s,a) + r\mathbb{E}_{a_i \sim \pi(a|s')}Q(s',a_i)$ with probability of $a_i \sim \pi(a|s')$ being $(1-\epsilon)$ optimal action and $\epsilon$ random action
  • update $Q(s,a) \leftarrow (1-\alpha)Q(s,a) + \alpha\hat{Q}(s,a)$
• compare the learning curve of Q-learning and SARSA

• implement experiment replay for off-line policy, Q-learning in this implementation

• Limitations of tabular methods
• Deep Q-network

• implement a neural network to approximate the action-value function $Q(s,a)$
• solving continuous state space environment 'CartPole-v0'

• implement DQN for 'BreakoutNoFrameskip-v4' environment
  • image processing
  • frame buffer
  • deep Q-network
  • experience replay
  • target networks

• Policy-based RL vs value-based RL
  • policy gradient
• REINFORCE
  • $\nabla_\theta\hat{J}(\theta) \approx \frac{1}{N}\Sigma_{s_i,a_i}\nabla_\theta \log\pi_\theta(a_i|s_i) \cdot G_t(s_i,a_i)$
• Actor-critic method
  • advantage actor-critic

• implement policy gradient method REINFORCE to solve 'CartPole-v0'

• Measuring the quality of exploration
  • Regret
• Uncertainty-based exploration
  • Thompson sampling: sample while taking into account actual distribution of rewards
  • Optimism in the face of uncertainty
  • UCB-1: select among actions that are uncertain or have potential to be optimal
  • Bayesian UCB
• Planning with Monte Carlo tree search
  • MCTS
    • selection: from root state recursively select node (next state by action) with exploration strategies (tree policy)
    • expansion: expand the node with feasible actions
    • simulation: estimate the value of the node by a rollout policy or some estimation model
    • backpropagation: propagate the estimation of action-state value to upper node utile root

• implement exploration strategies for bernoulli bandit
  • epsilon-greedy agent
  • UCB agent
  • Thompson sampling

• implement Monte Carlo Tree Search algorithm to plan and solve the 'CartPole-v0' task
  • selection (tree policy) is explored by UCB1
  • simulation is rollout by random selection of actions