# ECSE 506: Stochastic Control and Decision Theory

Example: Inventory Management (revisited)

TL;DR One of the potential benefits of modeling a system as infinite horizon discounted cost MDP is that it can be simpler to identify an optimal policy. We illustrate this using the inventory management example.

Consider the model for inventory management and assume that it runs for an infinite horizon. We assume that the per-step cost is given by $c(s,a,s_{+}) = p a + γ h(s),$ where $h(s) = \begin{cases} c_h s, & \text{if s \ge 0} \\ -c_s s, & \text{if s < 0}, \end{cases}$ where $$a$$ is the per-unit holding cost, $$b$$ is the per-unit shortage cost, and $$p$$ is the per-unit procurement cost. Note that in the per-step cost, we are assuming that the holding or shortage cost is dicounted because this cost is incurred at the end of the time period.

Recall that in the finite horizon setting, the optimal policy was a base-stock policy characterized by thresholds $$\{s^*_t\}_{t \ge 1}$$. In the infinite horizon discounted setting, we expect the optimal policy to be time-homogeneous, i.e., the thresholds $$\{s^*_t\}_{t \ge 1}$$ be a constant $$s^*$$ and not to depend on time.

As an illustration, let’s reconsider the example used for the finite horizon setting (where $$p=5$$, $$c_h = 4$$, $$c_s = 2$$, and the demand is Binomial(10,0.4)). We consider the discount factor $$γ = 0.9$$. The value function and optimal policy is this case are shown below.

We are interested in the following question: Is it possible to identify the optimal threshold of the base-stock policy without explicitly solving the dynamic program? In this section, we show that the answer is affirmative.

As a first step, we modify the per-step cost using reward shaping. In particular, we consider the following potential function

$\varphi(s) = h(s) + \frac1{γ} p s - \frac{1}{1-γ}p\mu,$ where $$\mu = \EXP[W]$$ is the expected number of arrivals at each time period.

Now consider a new cost function \begin{align*} c'(s,a,s_{+}) &= c(s,a,s_{+}) + \varphi(s) - γ \varphi(s_{+}) \\ &= pa + γ h(s_{+}) + h(s) + \frac{1}{γ} p s - \frac{1}{1-γ} p \mu - γ h(s_{+}) - p s_{+} - \frac{γ}{1-γ} p \mu \\ &= h(s) + \frac{1-γ}{γ} ps + p w - p \mu. \end{align*} Note that $\EXP[ c'(s,a,S_{+}) | S = s, A = a ] = h(s) + \frac{1-γ}{γ} ps =: c^*(s).$ Thus, the optimal policy of the original model is the same as that in which the per-step cost is given by $$c^*(s)$$.

Recall that the optimal policy in the original model was a base stock policy. For the infinite horizon model, the threshold will become time-invariant. Thus, the optimal policy will be of the form $π(s) = \begin{cases} s^* - s, & \text{if s \le s^*} \\ 0, & \text{otherwise}. \end{cases}$

The infinite horizon dynamic programming with this modified cost is given by $$$\label{eq:DP} V(s) = \min_{a \in \reals_{\ge 0}} \bigl\{ c^*(s) + γ \EXP[ V(s + a - W) ] \bigr\}.$$$

Using the structure of the optimal policy identified above, we have two properties. First, $$$\label{eq:opt} V(s) = c^*(s) + γ \EXP[ V(s^* - W) ], \qquad s \le s^*.$$$ Second, at $$s = 0$$, $$π(0) = s^*$$ (and recall that $$c^*(0) = 0$$). Therefore, $$$\label{eq:opt-policy} s^* = \arg\min_{a \in \reals_{\ge 0}} γ \EXP[V(a - W)]$$$

Let $$F(s^*)$$ denote $$\EXP[V(s^*-W)]$$. Then, substituting $$s = s^* - W$$ in \eqref{eq:opt} and taking expectations, we get $F(s^*) = \EXP[ c^*(s^* - W) ] + γ F(s^*).$ Thus, $\EXP[V(s^* - W)] = F(s^*) = \frac{1}{1-γ} \EXP[ c^*(s^*-W) ].$

Substituting the above in \eqref{eq:opt-policy}, we get $s^* = \arg\min_{s^* \ge 0} \frac{γ}{1-γ} \EXP[ c^*(s^* - W) ].$ Consequently, we have the following:

Theorem 1

The optimal threshold $$s^*$$ is given by the value of $$s^*$$ which minimizes $$\EXP[ c^*(s^*-W) ]$$. In particular, if $$F$$ denotes the CDF of the demand, then $$$s^* = F^{-1}\left( \frac{c_s - p(1-γ)/γ}{c_h + c_s} \right). \label{eq:opt-threshold}$$$

#### Proof

The proof idea is the same approach as that used for the newsvendor problem. In particular, let $$f$$ denote the distribution of $$W$$, $$F$$ denote the CDF, and $$μ = \EXP[W]$$. Then, \begin{align} \EXP[c^*(s^*-W)] &= \EXP[h(s^*-W)] + \frac{1-γ}{γ}p(s^* - μ) \notag \\ &= \int_{0}^{s^*} c_h(s^*-w)f(w)dw + \int_{s^*}^{∞} c_s(w-s^*)f(w)dw + \frac{1-γ}{γ}p(s^* - μ) \notag \end{align} Thereore, \begin{align} \frac{∂ \EXP[c^*(s^*-W)]}{∂s^*} &= \int_{0}^{s^*} c_h f(w)dw + \int_{s^*}^{\infty}[-c_s] f(w)dw + \frac{1-γ}{γ}p \notag \\ &= c_h F(s^*) - c_s(1 - F(s^*)) + \frac{1-γ}{γ}p \notag \\ \end{align} To find the optimal threshold, we set the derivative to $$0$$ and simplify to obtain \eqref{eq:opt-threshold}.

We can use \eqref{eq:opt-threshold} to find the optimal threshold for the example used above. In particular, we need to find the value of $$s^*$$ at which the CDF of the demand equals the value $$(c_s - p(1-γ)/γ)/(c_h + c_s)$$, as shown in Fig. 2 below.

# Exercises

1. Suppose that the arrival process is exponential with rate $$1/\mu$$, i.e., the density of $$W$$ is given by $$e^{-s/\mu}/\mu$$. Show that the optimal threshold is given by $s^* = \mu \log \left[ \frac{ c_h + c_s} { c_h + p (1-γ)/γ} \right].$

Hint: Recall that the CDF the exponential distribution is $$F(s) = 1 - e^{-s/μ}$$.

# References

The idea of using reward shaping to derive a closed form expression for inventory management is taken from Whittle (1982). It is interesting to note that Whittle (1982) uses the idea of reward shaping more than 17 years before the paper by Ng et al. (1999) on reward shaping. It is possible that Whittle was using the results of Porteus (1975).

Ng, A.Y., Harada, D., and Russell, S. 1999. Policy invariance under reward transformations: Theory and application to reward shaping. ICML, 278–287. Available at: http://aima.eecs.berkeley.edu/~russell/papers/icml99-shaping.pdf.
Porteus, E.L. 1975. Bounds and transformations for discounted finite markov decision chains. Operations Research 23, 4, 761–784. DOI: 10.1287/opre.23.4.761.
Whittle, P. 1982. Optimization over time: Dynamic programming and stochastic control. Vol. 1 and 2. Wiley.

This entry was last updated on 03 Oct 2022 and posted in MDP and tagged inventory management, base-stock policy, reward shaping, structural results, stochastic optimization, infinite horizon, discounted cost.