DP with Time Delays

Handling systems where the future depends on more than just the immediate past.

🐢

1. Introduction to Time Delays

Definition

In many applications, the next state \(x_{k+1}\) is influenced not just by \(x_k\) and \(u_k\), but also by previous states \(x_{k-1}, \dots\) and controls \(u_{k-1}, \dots\).

Consider a system with at most one stage delay:

\[ x_{k+1} = f_k(x_k, x_{k-1}, u_k, u_{k-1}) \]

System with Delay

graph LR xk["x_k"] --> System xk1["x_{k-1}"] --> System uk["u_k"] --> System uk1["u_{k-1}"] --> System System["System Dynamics f_k"] --> xk_new["x_{k+1}"] style System fill:#e0f2fe,stroke:#0284c7,stroke-width:2px
📦

2. State Augmentation

We can transform this into a standard system by introducing new state variables:

\[ y_k = x_{k-1}, \quad z_k = u_{k-1} \]

Augmented State Vector

Define the new state \(\tilde{x}_k = (x_k, y_k, z_k)\). The system becomes:

\[ \tilde{x}_{k+1} = \tilde{f}_k(\tilde{x}_k, u_k, w_k) \]

Augmented State Visualization

graph TD subgraph AS["Augmented State"] direction TB x["x_k (Current State)"] y["y_k (Previous State x_{k-1})"] z["z_k (Previous Control u_{k-1})"] end style AS fill:#f0fdf4,stroke:#16a34a,stroke-width:2px
🧮

3. DP Algorithm with Delays

By expressing the cost in terms of \(\tilde{x}_k\), we get a problem without delays. The policy \(\mu_k\) now depends on \((x_k, x_{k-1}, u_{k-1})\).

Reformulated Bellman Equation

\[ J^*_k(x_k, x_{k-1}, u_{k-1}) = \min_{u_k} E_{w_k} \left[ g_k + J^*_{k+1}(f_k(\dots), x_k, u_k) \right] \]
🔗

4. Non-additive Cost Structures

In extreme cases, cost might depend on the entire history:

\[ E\{g_N(x_N, \dots, x_0, u_{N-1}, \dots, u_0)\} \]

We augment the state to include the entire history:

\[ \tilde{x}_k = (x_k, \dots, x_0, u_{k-1}, \dots, u_0) \]
📊

5. Summary Infographic

DP with Time Delays Infographic
🧠

6. Test Your Knowledge

1. To handle a 1-step time delay in state \(x_{k-1}\), we augment the state with:

2. If the control \(u_{k-1}\) affects \(x_{k+1}\), the policy \(\mu_k\) must depend on:

3. State augmentation transforms a delayed system into:

4. For non-additive costs depending on full history, the state dimension:

5. Can we handle delays in disturbances \(w_{k-1}\)?

🔍 Spot the Mistake!

Scenario 1:

"We can ignore past controls \(u_{k-1}\) if only \(x_{k-1}\) affects the dynamics."

True/False? (Click to reveal: It depends on the specific dynamics f_k.)

Scenario 2:

"Augmenting the state reduces the computational complexity."

False. It drastically increases complexity.

Scenario 3:

"The cost function must always be additive for DP to work."

False. We can handle non-additive costs by augmenting the state.

Previous

Lecture 05

Next

Coming Soon