December 2018
Intermediate to advanced
764 pages
18h 18m
English
If the model is known beforehand then the agent can perform a policy search to find the optimal policy that maximizes the value function. When the model is available, the agent uses a value function that can be defined naively as a sum of the rewards of the future states:

Thus, the value at time-step t for selecting actions using the policy p would be:
V is the value and R is the reward, and the value function is estimated only up to n time-steps in the future.
When the agent estimates ...
Read now
Unlock full access