---

tags:

• 🌱
• AI
• ComputerScience
• MDP date: 19--Apr--2023

---

• Current rewards now are more valuable than rewards in the future
• Without discounting rewards, algorithm may run into infinite time horizon

• Discounting future rewards with $0\lt\gamma\lt1$
• $G_t=R_t+\gamma R_{t+1}+\gamma^2R_{t+2}+...=\sum_0^K{\gamma^i R_{t+i}}$
• $\displaystyle\sum_0^\infty{\gamma^i R_{t+i}}=\frac{R_{max}}{1-\gamma}$
• Bounds the accumulated rewards to ensure that algorithm converges

• $G_t=R_t+\gamma R_{t+1}+\gamma^2R_{t+2}+…$
• $G_t=R_t+\gamma G_{t+1}$

---

Links: