Deep reinforcement learning uses neural networks to approximate decision functions when an agent faces too many possible states for a table of values. A network can represent action values, a policy, or both. section 8.0 section 12.1
Algorithms differ in how they fill the shared structure:
- DQN learns action values and acts greedily; REINFORCE learns a policy directly; PPO combines a policy and critic; GRPO uses groups of sampled answers as a baseline. section 12.1
- To compare methods, look for which part changes while the broader learning loop remains. section 12.1
Sequential decisions
Cancer treatment can be framed as adjusting doses over multiple sessions as tumor response and side effects evolve. A portfolio agent can instead use a policy to choose among allocation options. section 8.0 section 12.1
Figure 12.1 The anatomy of a deep reinforcement learning algorithm.

Use the algorithm anatomy figure to orient the shared parts and loop. section 12.1

