What is deep reinforcement learning?
Programs that learn from trying, not from being told the answer.
In reinforcement learning (RL), an agent (the decision maker) tries actions in an environment and receives a reward in return. The goal is to find the behaviour that collects the most reward over the long run.
When multi-layer neural networks (deep learning) are used to approximate values, we call it deep RL.
Why RL is hard: three kinds of feedback at once
SequentialThe effect of an action may arrive much later. You must balance short- and long-term goals.
EvaluativeYou only get a score, never the correct action. You must balance gathering and using information.
SampledThere are too many states to see them all. You must generalise to new ones — which is where neural networks help.
Where RL sits in AI
| Branch | Learns from | Example |
|---|---|---|
| Supervised | Correct answers (labels) | Telling cats from dogs |
| Unsupervised | No labels; finds structure itself | Customer segmentation |
| Reinforcement | Rewards from trying | Walking robots, games, energy control |
Key takeaways
- DRL = trial and error + sequential, evaluative, sampled feedback + neural networks
- Main challenges: sample efficiency, good exploration and safety
- Handing important decisions to machines always calls for care
- #DeepRL
- #ReinforcementLearning
- #PythonສຳລັບຄົນLaos
- #AI
- #ບອກຕໍ່ຄວາມຮູ້