The Fascinating World of Machine Learning
Types of Machine Learning:
Machine learning comes in three flavors:
- Supervised Learning: This involves training a model on a labeled dataset.
- Unsupervised Learning: Here, the model is trained on an unlabeled dataset to recognize patterns.
- Reinforcement Learning: In this method, an agent learns through interaction and feedback from the environment.
Challenges:
One challenge in Reinforcement Learning is the credit assignment problem, which can be tackled with sparse rewards. Other challenges include overfitting reward functions and restricting the agent’s adaptability.
Efficient methods like adjusting pixel quality and predicting rewards can enhance sparse reward RL in games.
Proximal Policy Optimisation
Proximal Policy Optimisation is a policy gradient method that directly optimizes policies without storing previous results. It involves modeling the policy and optimizing it for the best reward.
It learns through trial and error by estimating rewards for actions in different states.
Q(s, a) = R(s, a) + γ * max(Q(s', a'))
Limitations of Q-Learning
- Requires a finite and discrete set of states and actions.
- As the action vector size grows, the space for the Q-table also increases.