Problems

1 / 1

Advantage Computation

Reinforcement Learning

Easy

Compute the Monte Carlo advantage at every time step of a completed episode. First compute each discounted return backward:

G_t=r_t+\gamma G_{t+1},\qquad G_T=0

Then subtract the value baseline for the visited state:

A_t=G_t-V(s_t)

Here, $r_t$ is the reward at time $t$ , $\gamma$ is the discount factor, and $V(s_t)$ is the estimated value of the current state. Do not bootstrap beyond the final reward. Return the advantages as a NumPy array rounded to four decimals.

Loading visualization...

Examples

Input: states = [0, 1, 2], rewards = [1, 2, 3], V = [0.5, 1, 1.5], gamma = 1

Output: [5.5, 4, 1.5]

Explanation: The returns are [6, 5, 3], and subtracting the three state baselines gives the advantages.

Input: states = [0, 1, 2], rewards = [1, 2, 3], V = [0, 0, 0], gamma = 0

Output: [1, 2, 3]

Hint 1

Fill a return array backward with running = rewards[index] + gamma * running.

Hint 2

Index the baseline vector with values[np.asarray(states, dtype=int)].

Requirements

Compute all returns in one backward pass
Subtract the value associated with each visited state
Do not add a value estimate after the final reward
Return a NumPy array rounded to four decimals

Constraints

states and rewards are nonempty lists of equal length
Every state index is valid for V
0 <= gamma <= 1
Use NumPy only

Try Similar Problems

Gae Computation Policy Gradient Loss Discount Returns Td Value Update Q Learning Update

Prev Next

Advantage Computation

Reinforcement Learning

Easy

Compute the Monte Carlo advantage at every time step of a completed episode. First compute each discounted return backward:

G_t=r_t+\gamma G_{t+1},\qquad G_T=0

Then subtract the value baseline for the visited state:

A_t=G_t-V(s_t)

Loading visualization...

Examples

Input: states = [0, 1, 2], rewards = [1, 2, 3], V = [0.5, 1, 1.5], gamma = 1

Output: [5.5, 4, 1.5]

Explanation: The returns are [6, 5, 3], and subtracting the three state baselines gives the advantages.

Input: states = [0, 1, 2], rewards = [1, 2, 3], V = [0, 0, 0], gamma = 0

Output: [1, 2, 3]

Hint 1

Fill a return array backward with running = rewards[index] + gamma * running.

Hint 2

Index the baseline vector with values[np.asarray(states, dtype=int)].

Requirements

Compute all returns in one backward pass
Subtract the value associated with each visited state
Do not add a value estimate after the final reward
Return a NumPy array rounded to four decimals

Constraints

states and rewards are nonempty lists of equal length
Every state index is valid for V
0 <= gamma <= 1
Use NumPy only

Try Similar Problems

Gae Computation Policy Gradient Loss Discount Returns Td Value Update Q Learning Update

Case 1

Case 2

states =

Accepts: array

rewards =

Accepts: array

V =

Accepts: array

gamma =

Accepts: number