julia
- depth
2
In this lecture we discuss a family of dynamic programming problems with the following features:
- a discrete state space and discrete choices (actions)
- an infinite horizon
- discounted rewards
- Markov state transitions
We call such problems discrete dynamic programs, or discrete DPs.
Discrete DPs are the workhorses in much of modern quantitative economics, including
- monetary economics
- search and labor economics
- household savings and consumption theory
- investment theory
- asset pricing
- industrial organization, etc.
When a given model is not inherently discrete, it is common to replace it with a discretized version in order to use discrete DP techniques.
This lecture covers
- the theory of dynamic programming in a discrete setting, plus examples and applications
- a powerful set of routines for solving discrete DPs from the QuantEcon code libary
We use dynamic programming many applied lectures, such as
- The
shortest path lecture <../dynamic_programming/short_path>
- The
McCall search model lecture <../dynamic_programming/mccall_model>
- The
optimal growth lecture <../dynamic_programming/optgrowth>
The objective of this lecture is to provide a more systematic and theoretical treatment, including algorithms and implementation, while focusing on the discrete case.
For background reading on dynamic programming and additional applications, see, for example,
Ljungqvist2012
HernandezLermaLasserre1996
, section 3.5puterman2005
StokeyLucas1989
Rust1996
MirandaFackler2002
- EDTC, chapter 5
Loosely speaking, a discrete DP is a maximization problem with an objective function of the form
where
- st is the state variable
- at is the action
- β is a discount factor
- r(st, at) is interpreted as a current reward when the state is st and the action chosen is at
Each pair (st, at) pins down transition probabilities Q(st, at, st + 1) for the next period state st + 1.
Thus, actions influence not only current rewards but also the future time path of the state.
The essence of dynamic programming problems is to trade off current rewards vs favorable positioning of the future state (modulo randomness).
Examples:
- consuming today vs saving and accumulating assets
- accepting a job offer today vs seeking a better one in the future
- exercising an option now vs waiting
The most fruitful way to think about solutions to discrete DP problems is to compare policies.
In general, a policy is a randomized map from past actions and states to current action.
In the setting formalized below, it suffices to consider so-called stationary Markov policies, which consider only the current state.
In particular, a stationary Markov policy is a map σ from states to actions
- at = σ(st) indicates that at is the action to be taken in state st
It is known that, for any arbitrary policy, there exists a stationary Markov policy that dominates it at least weakly.
- See section 5.5 of
puterman2005
for discussion and proofs.
In what follows, stationary Markov policies are referred to simply as policies.
The aim is to find an optimal policy, in the sense of one that maximizes dp_objective
.
Let's now step through these ideas more carefully.
Formally, a discrete dynamic program consists of the following components:
- A finite set of states S = {0, …, n − 1}
A finite set of feasible actions A(s) for each state s ∈ S, and a corresponding set of feasible state-action pairs
SA := {(s, a) ∣ s ∈ S, a ∈ A(s)}- A reward function r: SA → ℝ
- A transition probability function Q: SA → Δ(S), where Δ(S) is the set of probability distributions over S
- A discount factor β ∈ [0, 1)
We also use the notation A := ⋃s ∈ SA(s) = {0, …, m − 1} and call this set the action space.
A policy is a function σ: S → A.
A policy is called feasible if it satisfies σ(s) ∈ A(s) for all s ∈ S.
Denote the set of all feasible policies by Σ.
If a decision maker uses a policy σ ∈ Σ, then
- the current reward at time t is r(st, σ(st))
- the probability that st + 1 = s′ is Q(st, σ(st), s′)
For each σ ∈ Σ, define
- rσ by rσ(s) := r(s, σ(s)))
- Qσ by Qσ(s, s′) := Q(s, σ(s), s′)
Notice that Qσ is a stochastic matrix <finite_dp_stoch_mat>
on S.
It gives transition probabilities of the controlled chain when we follow policy σ.
If we think of rσ as a column vector, then so is Qσtrσ, and the s-th row of the latter has the interpretation
(Qσtrσ)(s) = 𝔼[r(st, σ(st)) ∣ s0 = s] when {st} ∼ Qσ
Comments
- {st} ∼ Qσ means that the state is generated by stochastic matrix Qσ
- See
this discussion <finite_mc_expec>
on computing expectations of Markov chains for an explanation of the expression inddp_expec
Notice that we're not really distinguishing between functions from S to ℝ and vectors in ℝn.
This is natural because they are in one to one correspondence.
Let vσ(s) denote the discounted sum of expected reward flows from policy σ when the initial state is s.
To calculate this quantity we pass the expectation through the sum in dp_objective
and use ddp_expec
to get
This function is called the policy value function for the policy σ.
The optimal value function, or simply value function, is the function v*: S → ℝ defined by
v*(s) = maxσ ∈ Σvσ(s) (s ∈ S)
(We can use max rather than sup here because the domain is a finite set)
A policy σ ∈ Σ is called optimal if vσ(s) = v*(s) for all s ∈ S.
Given any w: S → ℝ, a policy σ ∈ Σ is called w-greedy if
As discussed in detail below, optimal policies are precisely those that are v*-greedy.
It is useful to define the following operators:
- The Bellman operator T: ℝS → ℝS is defined by
(Tv)(s) = maxa ∈ A(s){r(s,a)+β∑s′ ∈ Sv(s′)Q(s,a,s′)} (s ∈ S)
- For any policy function σ ∈ Σ, the operator Tσ: ℝS → ℝS is defined by
(Tσv)(s) = r(s, σ(s)) + β∑s′ ∈ Sv(s′)Q(s, σ(s), s′) (s ∈ S)
This can be written more succinctly in operator notation as
Tσv = rσ + βQσv
The two operators are both monotone
- v ≤ w implies Tv ≤ Tw pointwise on S, and similarly for Tσ
They are also contraction mappings with modulus β
- ∥Tv − Tw∥ ≤ β∥v − w∥ and similarly for Tσ, where ∥ ⋅ ∥ is the max norm
For any policy σ, its value vσ is the unique fixed point of Tσ.
For proofs of these results and those in the next section, see, for example, EDTC, chapter 10.
The main principle of the theory of dynamic programming is that
the optimal value function v* is a unique solution to the Bellman equation,
v(s) = maxa ∈ A(s){r(s,a)+β∑s′ ∈ Sv(s′)Q(s,a,s′)} (s ∈ S),or in other words, v* is the unique fixed point of T, and
- σ* is an optimal policy function if and only if it is v*-greedy
By the definition of greedy policies given above, this means that
Now that the theory has been set out, let's turn to solution methods.
Code for solving discrete DPs is available in ddp.jl from the QuantEcon.jl code library.
It implements the three most important solution methods for discrete dynamic programs, namely
- value function iteration
- policy function iteration
- modified policy function iteration
Let's briefly review these algorithms and their implementation.
Perhaps the most familiar method for solving all manner of dynamic programs is value function iteration.
This algorithm uses the fact that the Bellman operator T is a contraction mapping with fixed point v*.
Hence, iterative application of T to any initial function v0: S → ℝ converges to v*.
The details of the algorithm can be found in the appendix <ddp_algorithms>
.
This routine, also known as Howard's policy improvement algorithm, exploits more closely the particular structure of a discrete DP problem.
Each iteration consists of
- A policy evaluation step that computes the value vσ of a policy σ by solving the linear equation v = Tσv.
- A policy improvement step that computes a vσ-greedy policy.
In the current setting policy iteration computes an exact optimal policy in finitely many iterations.
- See theorem 10.2.6 of EDTC for a proof
The details of the algorithm can be found in the appendix <ddp_algorithms>
.
Modified policy iteration replaces the policy evaluation step in policy iteration with "partial policy evaluation".
The latter computes an approximation to the value of a policy σ by iterating Tσ for a specified number of times.
This approach can be useful when the state space is very large and the linear system in the policy evaluation step of policy iteration is correspondingly difficult to solve.
The details of the algorithm can be found in the appendix <ddp_algorithms>
.
Let's consider a simple consumption-saving model.
A single household either consumes or stores its own output of a single consumption good.
The household starts each period with current stock s.
Next, the household chooses a quantity a to store and consumes c = s − a
- Storage is limited by a global upper bound M
- Flow utility is u(c) = cα
Output is drawn from a discrete uniform distribution on {0, …, B}.
The next period stock is therefore
s′ = a + U where U ∼ U[0, …, B]
The discount factor is β ∈ [0, 1).
We want to represent this model in the format of a discrete dynamic program.
To this end, we take
- the state variable to be the stock s
the state space to be S = {0, …, M + B}
- hence n = M + B + 1
- the action to be the storage quantity a
the set of feasible actions at s to be A(s) = {0, …, min {s, M}}
- hence A = {0, …, M} and m = M + 1
- the reward function to be r(s, a) = u(s − a)
- the transition probabilities to be
This information will be used to create an instance of DiscreteDP by passing the following information
- An n × m reward array R
- An n × m × n transition probability array Q
- A discount factor β
For R we set R[s, a] = u(s − a) if a ≤ s and − ∞ otherwise.
For Q we follow the rule in ddp_def_ogq
.
Note:
- The feasibility constraint is embedded into R by setting R[s, a] = − ∞ for a ∉ A(s).
- Probability distributions for (s, a) with a ∉ A(s) can be arbitrary.
The following code sets up these objects for us.
/_static/includes/deps_generic.jl
using BenchmarkTools, Plots, QuantEcon, Parameters
gr(fmt = :png);
SimpleOG = @with_kw (B = 10, M = 5, α = 0.5, β = 0.9)
function transition_matrices(g)
@unpack B, M, α, β = g
u(c) = c^α
n = B + M + 1
m = M + 1
R = zeros(n, m)
Q = zeros(n, m, n)
for a in 0:M
Q[:, a + 1, (a:(a + B)) .+ 1] .= 1 / (B + 1)
for s in 0:(B + M)
R[s + 1, a + 1] = (a≤s ? u(s - a) : -Inf)
end
end
return (Q = Q, R = R)
end
Let's run this code and create an instance of SimpleOG
g = SimpleOG();
Q, R = transition_matrices(g);
In case the preceding code was too concise, we can see a more verbose form
function verbose_matrices(g)
@unpack B, M, α, β = g
u(c) = c^α
#Matrix dimensions. The +1 is due to the 0 state.
n = B + M + 1
m = M + 1
R = fill(-Inf, n, m) #Start assuming nothing is feasible
Q = zeros(n,m,n) #Assume 0 by default
#Create the R matrix
#Note: indexing into matrix complicated since Julia starts indexing at 1 instead of 0
#but the state s and choice a can be 0
for a in 0:M
for s in 0:(B + M)
if a <= s #i.e. if feasible
R[s + 1, a + 1] = u(s - a)
end
end
end
#Create the Q multi-array
for s in 0:(B+M) #For each state
for a in 0:M #For each action
for sp in 0:(B+M) #For each state next period
if( sp >= a && sp <= a + B) # The support of all realizations
Q[s + 1, a + 1, sp + 1] = 1 / (B + 1) # Same prob of all
end
end
@assert sum(Q[s + 1, a + 1, :]) ≈ 1 #Optional check that matrix is stochastic
end
end
return (Q = Q, R = R)
end
Instances of DiscreteDP
are created using the signature DiscreteDP(R, Q, β)
.
Let's create an instance using the objects stored in g
ddp = DiscreteDP(R, Q, g.β);
Now that we have an instance ddp
of DiscreteDP
we can solve it as follows
results = solve(ddp, PFI)
Let's see what we've got here
fieldnames(typeof(results))
The most important attributes are v
, the value function, and σ
, the optimal policy
results.v
results.sigma .- 1
Here 1 is subtracted from results.sigma because we added 1 to each state and action to create valid indices.
Since we've used policy iteration, these results will be exact unless we hit the iteration bound max_iter
.
Let's make sure this didn't happen
results.num_iter
In this case we converged in only 3 iterations.
Another interesting object is results.mc
, which is the controlled chain defined by Qσ*, where σ* is the optimal policy.
In other words, it gives the dynamics of the state when the agent follows the optimal policy.
Since this object is an instance of MarkovChain from QuantEcon.jl (see this lecture <../tools_and_techniques/finite_markov>
for more discussion), we can easily simulate it, compute its stationary distribution and so on
stationary_distributions(results.mc)[1]
Here's the same information in a bar graph
What happens if the agent is more patient?
g_2 = SimpleOG(β=0.99);
Q_2, R_2 = transition_matrices(g_2);
ddp_2 = DiscreteDP(R_2, Q_2, g_2.β)
results_2 = solve(ddp_2, PFI)
std_2 = stationary_distributions(results_2.mc)[1]
bar(std_2, label = "stationary dist")
We can see the rightward shift in probability mass.
The DiscreteDP
type in fact provides a second interface to setting up an instance.
One of the advantages of this alternative set up is that it permits use of a sparse matrix for Q
.
(An example of using sparse matrices is given in the exercises below)
The call signature of the second formulation is DiscreteDP(R, Q, β, s_indices, a_indices)
where
s_indices
anda_indices
are arrays of equal lengthL
enumerating all feasible state-action pairsR
is an array of lengthL
giving corresponding rewardsQ
is anL x n
transition probability array
Here's how we could set up these objects for the preceding example
B = 10
M = 5
α = 0.5
β = 0.9
u(c) = c^α
n = B + M + 1
m = M + 1
s_indices = Int64[]
a_indices = Int64[]
Q = zeros(0, n)
R = zeros(0)
b = 1 / (B + 1)
for s in 0:(M + B)
for a in 0:min(M, s)
s_indices = [s_indices; s + 1]
a_indices = [a_indices; a + 1]
q = zeros(1, n)
q[(a + 1):((a + B) + 1)] .= b
Q = [Q; q]
R = [R; u(s-a)]
end
end
ddp = DiscreteDP(R, Q, β, s_indices, a_indices);
results = solve(ddp, PFI)
In the stochastic optimal growth lecture dynamic programming lecture <../dynamic_programming/optgrowth>
, we solve a benchmark model <benchmark_growth_mod>
that has an analytical solution to check we could replicate it numerically.
The exercise is to replicate this solution using DiscreteDP
.
These were written jointly by Max Huber and Daisuke Oyama.
Details of the model can be found in the lecture. As in the lecture, we let f(k) = kα with α = 0.65, u(c) = log c, and β = 0.95.
α = 0.65
f(k) = k.^α
u_log(x) = log(x)
β = 0.95
Here we want to solve a finite state version of the continuous state model above. We discretize the state space into a grid of size grid_size = 500
, from 10 − 6 to grid_max=2
.
grid_max = 2
grid_size = 500
grid = range(1e-6, grid_max, length = grid_size)
We choose the action to be the amount of capital to save for the next period (the state is the capital stock at the beginning of the period). Thus the state indices and the action indices are both 1
, ..., grid_size
. Action (indexed by) a
is feasible at state (indexed by) s
if and only if grid[a] < f([grid[s])
(zero consumption is not allowed because of the log utility).
Thus the Bellman equation is:
v(k) = max0 < k′ < f(k)u(f(k) − k′) + βv(k′),
where k′ is the capital stock in the next period.
The transition probability array Q
will be highly sparse (in fact it is degenerate as the model is deterministic), so we formulate the problem with state-action pairs, to represent Q
in sparse matrix format.
We first construct indices for state-action pairs:
C = f.(grid) .- grid'
coord = repeat(collect(1:grid_size), 1, grid_size) #coordinate matrix
s_indices = coord[C .> 0]
a_indices = transpose(coord)[C .> 0]
L = length(a_indices)
Now let's set up R and Q
R = u_log.(C[C.>0]);
using SparseArrays
Q = spzeros(L, grid_size) # Formerly spzeros
for i in 1:L
Q[i, a_indices[i]] = 1
end
We're now in a position to create an instance of DiscreteDP
corresponding to the growth model.
ddp = DiscreteDP(R, Q, β, s_indices, a_indices);
results = solve(ddp, PFI)
v, σ, num_iter = results.v, results.sigma, results.num_iter
num_iter
Let us compare the solution of the discrete model with the exact solution of the original continuous model. Here's the exact solution:
c = f(grid) - grid[σ]
ab = α * β
c1 = (log(1 - α * β) + log(α * β) * α * β / (1 - α * β)) / (1 - β)
c2 = α / (1 - α * β)
v_star(k) = c1 + c2 * log(k)
c_star(k) = (1 - α * β) * k.^α
Let's plot the value functions.
plot(grid, [v v_star.(grid)], ylim = (-40, -32), lw = 2, label = ["discrete" "continuous"])
They are barely distinguishable (although you can see the difference if you zoom).
Now let's look at the discrete and exact policy functions for consumption.
plot(grid, [c c_star.(grid)], lw = 2, label = ["discrete" "continuous"], legend = :topleft)
These functions are again close, although some difference is visible and becomes more obvious as you zoom. Here are some statistics:
maximum(abs(x - v_star(y)) for (x, y) in zip(v, grid))
This is a big error, but most of the error occurs at the lowest gridpoint. Otherwise the fit is reasonable:
maximum(abs(v[idx] - v_star(grid[idx])) for idx in 2:lastindex(v))
The value function is monotone, as expected:
all(x -> x ≥ 0, diff(v))
Let's try different solution methods. The results below show that policy function iteration and modified policy function iteration are much faster that value function iteration.
@benchmark results = solve(ddp, PFI)
results = solve(ddp, PFI);
@benchmark res1 = solve(ddp, VFI, max_iter = 500, epsilon = 1e-4)
res1 = solve(ddp, VFI, max_iter = 500, epsilon = 1e-4);
res1.num_iter
σ == res1.sigma
@benchmark res2 = solve(ddp, MPFI, max_iter = 500, epsilon = 1e-4)
res2 = solve(ddp, MPFI, max_iter = 500, epsilon = 1e-4);
res2.num_iter
σ == res2.sigma
Let's visualize convergence of value function iteration, as in the lecture.
w_init = 5log.(grid) .- 25 # Initial condition
n = 50
ws = []
colors = []
w = w_init
for i in 0:n-1
w = bellman_operator(ddp, w)
push!(ws, w)
push!(colors, RGBA(0, 0, 0, i/n))
end
plot(grid,
w_init,
ylims = (-40, -20),
lw = 2,
xlims = extrema(grid),
label = "initial condition")
plot!(grid, ws, label = "", color = reshape(colors, 1, length(colors)), lw = 2)
plot!(grid, v_star.(grid), label = "true value function", color = :red, lw = 2)
We next plot the consumption policies along the value iteration. First we write a function to generate the and record the policies at given stages of iteration.
function compute_policies(n_vals...)
c_policies = []
w = w_init
for n in 1:maximum(n_vals)
w = bellman_operator(ddp, w)
if n in n_vals
σ = compute_greedy(ddp, w)
c_policy = f(grid) - grid[σ]
push!(c_policies, c_policy)
end
end
return c_policies
end
Now let's generate the plots.
true_c = c_star.(grid)
c_policies = compute_policies(2, 4, 6)
plot_vecs = [c_policies[1] c_policies[2] c_policies[3] true_c true_c true_c]
l1 = "approximate optimal policy"
l2 = "optimal consumption policy"
labels = [l1 l1 l1 l2 l2 l2]
plot(grid,
plot_vecs,
xlim = (0, 2),
ylim = (0, 1),
layout = (3, 1),
lw = 2,
label = labels,
size = (600, 800),
title = ["2 iterations" "4 iterations" "6 iterations"])
Finally, let us work on Exercise 2, where we plot the trajectories of the capital stock for three different discount factors, 0.9, 0.94, and 0.98, with initial condition k0 = 0.1.
discount_factors = (0.9, 0.94, 0.98)
k_init = 0.1
k_init_ind = findfirst(collect(grid) .≥ k_init)
sample_size = 25
ddp0 = DiscreteDP(R, Q, β, s_indices, a_indices)
k_paths = []
labels = []
for β in discount_factors
ddp0.beta = β
res0 = solve(ddp0, PFI)
k_path_ind = simulate(res0.mc, sample_size, init=k_init_ind)
k_path = grid[k_path_ind.+1]
push!(k_paths, k_path)
push!(labels, "β = $β")
end
plot(k_paths,
xlabel = "time",
ylabel = "capital",
ylim = (0.1, 0.3),
lw = 2,
markershape = :circle,
label = reshape(labels, 1, length(labels)))
This appendix covers the details of the solution algorithms implemented for DiscreteDP
.
We will make use of the following notions of approximate optimality:
- For ε > 0, v is called an ε-approximation of v* if ∥v − v*∥ < ε
- A policy σ ∈ Σ is called ε-optimal if vσ is an ε-approximation of v*
The DiscreteDP
value iteration method implements value function iteration as follows
- Choose any v0 ∈ ℝn, and specify ε > 0; set i = 0.
- Compute vi + 1 = Tvi.
- If ∥vi + 1 − vi∥ < [(1 − β)/(2β)]ε, then go to step 4; otherwise, set i = i + 1 and go to step 2.
- Compute a vi + 1-greedy policy σ, and return vi + 1 and σ.
Given ε > 0, the value iteration algorithm
- terminates in a finite number of iterations
- returns an ε/2-approximation of the optimal value function and an ε-optimal policy function (unless
iter_max
is reached)
(While not explicit, in the actual implementation each algorithm is terminated if the number of iterations reaches iter_max
)
The DiscreteDP
policy iteration method runs as follows
- Choose any v0 ∈ ℝn and compute a v0-greedy policy σ0; set i = 0.
- Compute the value vσi by solving the equation v = Tσiv.
- Compute a vσi-greedy policy σi + 1; let σi + 1 = σi if possible.
- If σi + 1 = σi, then return vσi and σi + 1; otherwise, set i = i + 1 and go to step 2.
The policy iteration algorithm terminates in a finite number of iterations.
It returns an optimal value function and an optimal policy function (unless iter_max
is reached).
The DiscreteDP
modified policy iteration method runs as follows:
- Choose any v0 ∈ ℝn, and specify ε > 0 and k ≥ 0; set i = 0.
- Compute a vi-greedy policy σi + 1; let σi + 1 = σi if possible (for i ≥ 1).
- Compute u = Tvi ( = Tσi + 1vi). If span(u − vi) < [(1 − β)/β]ε, then go to step 5; otherwise go to step 4.
- Span is defined by span(z) = max (z) − min (z)
- Compute vi + 1 = (Tσi + 1)ku ( = (Tσi + 1)k + 1vi); set i = i + 1 and go to step 2.
- Return v = u + [β/(1 − β)][(min (u − vi) + max (u − vi))/2]1 and σi + 1.
Given ε > 0, provided that v0 is such that Tv0 ≥ v0, the modified policy iteration algorithm terminates in a finite number of iterations.
It returns an ε/2-approximation of the optimal value function and an ε-optimal policy function (unless iter_max
is reached).
See also the documentation for DiscreteDP
.