Reinforcement Learning Playground

What is reinforcement learning?

• Reinforcement learning is a branch of machine learning.
• Involves an agent and environment.
• Agents learns optimal for maximizing rewards.

When should we worry about sequential decision making?

Limited supervision: you know what you want, but not how to get it.

Late consequences?

Why learn RL?

• Not just for games
• Make optimal decisions
• Maximize efficiency

What are RL applications?

• Robotics
• Self-driving cars
• Inventory management
• Finantial investments
• Decision-based situations

RL terminology

What is the agent?

• The agent is the algorithm
• Decides which action to tale
• Agent monitors the environment
• Who is learning
• It's only outcome are decisions(actions, controls)

What is an environment?

• The environment is everything the agent can interact with.
• Agent's actions affect the environment.
• It responds to actor's actions with consequences(observations, rewards estimation)

What is a state?

• The state is a representation of what the agent can sense.
• Does not always involve the entire environment. It's limited to what the agent can sense.

What is an action?

• An action is what an agent can do is a given state.
• Actions are limited by the environment.
• The action's goal is to maximize reward.

What is the reward?

• Result from making an action.
• Feedback from the environment.
• It can be positive or negative.
• Helps encourage or discourage certain actions, policies or behaivours.
• Is what the agent tries to optimize.
• Rewards are hard to formulate.

Where do rewards come from?

• When playing video games, rewards come from scores.

Are there other forms of supervision?

• Learning from demostrations.

  • Directly copying observed behavior.
  • Inferring rewards from observed behavior.

• Learning from observing the world.

  • Learning to predict.
  • Unsupervised Learning

• Learning from other tasks

  • Transfer learning

What is the standard reinforcement loop?

• TODO

What is Deep Reinforcement Learning?

• Deep learning: end-to-end training of expressive, multi-layer models.
• Deep models are what allow RL algorithms to solve complex problems end-to-end.

Why Deep Reinforcement Learning?

• Deep = can process complex sensory input

What can deep learning & RL do well now?

• Adquire high degree of proficiency in domains governed by simple, known rules.
• Learn simple skills with raw sensory inputs, given enough experience.
• Learn from imitating enough human-provided expert behavior.

What has proven challenging so far?

• Humans can learn incredibly quickly
• Humans can reuse past knowledge
  • Transfer learning in deep RL is an open problem
• Not clear what the reward function should be

How do we build intelligent machines?

Learning as the basis of intelligence.

• Some things we can all do.
• Some things we can only learn.