This question relates to this paper on Trust Region Policy Optimization.

I understand the reasoning behind the proof of Theorem 1 presented in Appendix A, but fail to see how the absolute value was removed in going from equation (45) to Theorem 1:

$$ |\eta(\tilde{\pi}) - L_{\pi}(\tilde{\pi})| \le \frac{4\alpha^2 \gamma \epsilon}{(1 - \gamma)^2} \tag{45}$$

$$ \eta(\tilde{\pi}) \ge L_{\pi}(\tilde{\pi}) - \frac{4\alpha^2 \gamma \epsilon}{(1 - \gamma)^2} \tag{Theorem 1}$$

This suggests that the approximation is an overestimate ($\eta(\tilde{\pi}) \le L_{\pi}(\tilde{\pi})$), but I don't see why that must be true.


If the approximation is an underestimate: $\eta(\tilde \pi) > L_\pi(\tilde \pi)$ and the RHS of (45) is non-negative, then the theorem is trivially true.

  • $\begingroup$ I see. It was a simple leap of logic that I missed. Specifically, if $\eta(\tilde{\pi}) > L_{\pi}(\tilde{\pi})$, then $\eta(\tilde{\pi}) - L_{\pi}(\tilde{\pi}) > 0 \ge -\frac{4\alpha^2 \gamma \epsilon}{(1 - \gamma)^2}$. $\endgroup$ – rish987 Jul 10 '18 at 14:02

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.