Model-Based Reinforcement Learning

Model-Based RL

only differs from its model-free counterpart in learning or using a dynamic model to improve policy learning. It learns/uses the model to predict the next state and reward given a state an action: $s_{t+1},r_{t+1}=f_{\theta }(s_t,a_t)$.

doesn’t require real-world interaction.

Advantages

• Reduces the need for real-world interactions by using synthetic data.
• Better exploration since the models predict outcomes of unexplored actions.
• Transfer Learning

Disadvantages

• Training models and planning can be expensive.
• Some methods struggle with high-dimensional or continuous spaces.