The systems & control research community has developed a range of tools for understanding and controlling complex systems. Some of these techniques are model-based: Using a simple model we obtain insight regarding the structure of effective policies for control. The talk will survey how this point of view can be applied to approach resource allocation problems, such as those that will arise in the next-generation energy grid. We also show how insight from this kind of analysis can be used to construct architectures for reinforcement learning algorithms used in a broad range of applications.
Much of the talk is a survey from a recent book by the author with a similar title,
Control Techniques for Complex Networks. Cambridge University Press, 2007.
https://netfiles.uiuc.edu/meyn/www/spm_files/CTCN/CTCN.html
1. Control Techniques for Complex Systems
Department of Electrical & Computer Engineering
University of Florida
Sean P. Meyn
Coordinated Science Laboratory
and the Department of Electrical and Computer Engineering
University of Illinois at Urbana-Champaign, USA
April 21, 2011
1 / 26
2. Outline
Control Techniques Markov Chains
FOR and
Complex Networks Stochastic Stability
P n (x, · ) − π f →0
sup Ex [SτC (f )] < ∞
C
π(f ) < ∞
1 Control Techniques ∆V (x) ≤ −f (x) + bIC (x)
Sean Meyn S. P. Meyn and R. L. Tweedie
2 Complex Networks
3 Architectures for Adaptation & Learning
4 Next Steps
2 / 26
3. Control Techniques
System model
d
α = µ σ −Cα + . . .
dt
d
q = 1 µ I −1 (C − . . .
2
dt
d
θ=q
dt
???
Control Techniques?
3 / 26
4. Control Techniques
Typical steps to control design
Obtain simple model that captures System model
essential structure d
dt
α = µ σ −Cα + . . .
d
– An equilibrium model if the goal is regulation dt
q = 1 µ I −1 (C − . . .
2
d
θ=q
dt
???
4 / 26
5. Control Techniques
Typical steps to control design
Obtain simple model that captures System model
essential structure d
dt
α = µ σ −Cα + . . .
d
– An equilibrium model if the goal is regulation dt
q = 1 µ I −1 (C − . . .
2
d
θ=q
dt
???
Obtain feedback design, using dynamic programming, LQG, loop shaping, ...
Design for performance and reliability
Test via simulations and experiments, and refine design
4 / 26
6. Control Techniques
Typical steps to control design
Obtain simple model that captures System model
essential structure d
dt
α = µ σ −Cα + . . .
d
– An equilibrium model if the goal is regulation dt
q = 1 µ I −1 (C − . . .
2
d
θ=q
dt
???
Obtain feedback design, using dynamic programming, LQG, loop shaping, ...
Design for performance and reliability
Test via simulations and experiments, and refine design
If these steps fail, we may have to re-engineer the
system (e.g., introduce new sensors), and start over.
4 / 26
7. Control Techniques
Typical steps to control design
Obtain simple model that captures System model
essential structure d
dt
α = µ σ −Cα + . . .
d
– An equilibrium model if the goal is regulation dt
q = 1 µ I −1 (C − . . .
2
d
θ=q
dt
???
Obtain feedback design, using dynamic programming, LQG, loop shaping, ...
Design for performance and reliability
Test via simulations and experiments, and refine design
If these steps fail, we may have to re-engineer the
system (e.g., introduce new sensors), and start over.
This point of view is unique to control
4 / 26
8. Control Techniques
Typical steps to scheduling
Inventory model: Controlled work-release, controlled routing,
uncertain demand
A simplified model of a semiconductor
manufacturing facility
Similar demand-driven models can be used
demand 1
to model allocation of locational reserves
in a power grid
demand 2
5 / 26
9. Control Techniques
Typical steps to scheduling
Inventory model: Controlled work-release, controlled routing,
uncertain demand
A simplified model of a semiconductor
manufacturing facility
Similar demand-driven models can be used
demand 1
to model allocation of locational reserves
in a power grid
demand 2
Obtain simple model –
Frequently based on simple statistics to obtain a Markov model
Obtain feedback design based on heuristics, or dynamic programming
Performance evaluation via computation
(e.g., Neuts’ matrix-geometric methods)
5 / 26
10. Control Techniques
Typical steps to scheduling
Inventory model: Controlled work-release, controlled routing,
uncertain demand
A simplified model of a semiconductor
manufacturing facility.
Similar demand-driven models can be used demand 1
to model allocation of locational reserves
in a power grid demand 2
Difficulty : A Markov model is not simple enough!
Obtain simple model –
Frequently based on exponential statistics to obtain a Markov model
Obtain feedback design based on heuristics, or dynamic programming
Performance evaluation via computation (e.g., Neut’s matrix-geometric methods)
With the 16 buffers truncated to 0 ≤ x ≤ 10,
6 / 26
11. Control Techniques
Typical steps to scheduling
Inventory model: Controlled work-release, controlled routing,
uncertain demand
A simplified model of a semiconductor
manufacturing facility.
Similar demand-driven models can be used demand 1
to model allocation of locational reserves
in a power grid demand 2
Difficulty : A Markov model is not simple enough!
Obtain simple model –
Frequently based on exponential statistics to obtain a Markov model
Obtain feedback design based on heuristics, or dynamic programming
Performance evaluation via computation (e.g., Neut’s matrix-geometric methods)
With the 16 buffers truncated to 0 ≤ x ≤ 10,
policy synthesis reduces to a linear program of dimension 1116 !
6 / 26
12. Control Techniques
Control-theoretic approach to scheduling d
dt q = Bu + α
Inventory model: Controlled work-release, controlled routing,
uncertain demand
q: Queue length evolves on R16 .
+
u: Scheduling/routing decisions —
demand 1
Convex relaxation
demand 2
α: Mean exogenous arrivals of work
B: Captures network topology
7 / 26
13. Control Techniques
Control-theoretic approach to scheduling d
dt q = Bu + α
Inventory model: Controlled work-release, controlled routing,
uncertain demand
q: Queue length evolves on R16 .
+
u: Scheduling/routing decisions —
demand 1
Convex relaxation
demand 2
α: Mean exogenous arrivals of work
B: Captures network topology
Control-theoretic approach to scheduling:
Dimension reduced from a linear program of dimension 1116 ...
to an HJB equation of dimension 16
7 / 26
14. Control Techniques
Control-theoretic approach to scheduling d
dt q = Bu + α
Inventory model: Controlled work-release, controlled routing,
uncertain demand
q: Queue length evolves on R16 .
+
u: Scheduling/routing decisions —
demand 1
Convex relaxation
demand 2
α: Mean exogenous arrivals of work
B: Captures network topology
Control-theoretic approach to scheduling:
Dimension reduced from a linear program of dimension 1116 ...
to an HJB equation of dimension 16
Does this solve the problem?
7 / 26
18. Complex Networks
Dynamic Programming Equations
Deterministic model x = f (x, u)
˙
Controlled generator
d
Du h (x) = dt h(x(t)) t=0
x(0)=x
u(0)=u
9 / 26
19. Complex Networks
Dynamic Programming Equations
Deterministic model x = f (x, u)
˙
Controlled generator
d
Du h (x) = dt h(x(t)) t=0 = f (x, u) · h (x)
x(0)=x
u(0)=u
9 / 26
20. Complex Networks
Dynamic Programming Equations
Deterministic model x = f (x, u)
˙
Controlled generator
d
Du h (x) = dt h(x(t)) t=0 = f (x, u) · h (x)
x(0)=x
u(0)=u
Minimal total cost:
∞
J ∗ (x) = inf c(x(t), u(t)) dt , x(0) = x
U 0
HJB Equation:
min c(x, u) + Du J ∗ (x) = 0
u
9 / 26
21. Complex Networks
Dynamic Programming Equations
Diffusion model dX = f (X, U )dt + σ(X)dN
Controlled generator
d
Du h (x) = E[h(X(t))] t=0
dt x(0)=x
u(0)=u
2
= f (x, u) · h (x) + 1 trace σ(x)T
2 h (x)σ(x)
10 / 26
22. Complex Networks
Dynamic Programming Equations
Diffusion model dX = f (X, U )dt + σ(X)dN
Controlled generator
d
Du h (x) = E[h(X(t))] t=0
dt x(0)=x
u(0)=u
2
= f (x, u) · h (x) + 1 trace σ(x)T
2 h (x)σ(x)
Minimal average cost:
T
1
η ∗ = inf lim c(X(t), U (t)) dt
U T →∞ T 0
10 / 26
23. Complex Networks
Dynamic Programming Equations
Diffusion model dX = f (X, U )dt + σ(X)dN
Controlled generator
d
Du h (x) = E[h(X(t))] t=0
dt x(0)=x
u(0)=u
2
= f (x, u) · h (x) + 1 trace σ(x)T
2 h (x)σ(x)
Minimal average cost:
T
1
η ∗ = inf lim c(X(t), U (t)) dt
U T →∞ T 0
ACOE (Average Cost Optimality Equation):
min c(x, u) + Du h∗ (x) = η ∗
u
h∗ is the relative value function
10 / 26
24. Complex Networks
Dynamic Programming Equations
MDP model X(t + 1) − X(t) = f (X(t), U (t), N (t + 1))
Controlled generator
Du h (x) = E[h(X(1)) − h(X(0))]
= E[h(x + f (x, u, N ))] − h(x)
11 / 26
25. Complex Networks
Dynamic Programming Equations
MDP model X(t + 1) − X(t) = f (X(t), U (t), N (t + 1))
Controlled generator
Du h (x) = E[h(X(1)) − h(X(0))]
= E[h(x + f (x, u, N ))] − h(x)
Minimal average cost:
T −1
∗ 1
η = inf lim c(X(t), U (t))
U T →∞ T
0
ACOE (Average Cost Optimality Equation):
min c(x, u) + Du h∗ (x) = η ∗
u
h∗ is the relative value function
11 / 26
26. Complex Networks
Approximate Dynamic Programming
ODE model from the MDP model, X(t + 1) − X(t) = f (X(t), U (t), N (t + 1))
Mean drift: f (x, u) = E[X(t + 1) − X(t) | X(t) = x, U (t) = u]
12 / 26
27. Complex Networks
Approximate Dynamic Programming
ODE model from the MDP model, X(t + 1) − X(t) = f (X(t), U (t), N (t + 1))
Mean drift: f (x, u) = E[X(t + 1) − X(t) | X(t) = x, U (t) = u]
Fluid Model: x(t) = f (x(t), u(t))
˙
12 / 26
28. Complex Networks
Approximate Dynamic Programming
ODE model from the MDP model, X(t + 1) − X(t) = f (X(t), U (t), N (t + 1))
Mean drift: f (x, u) = E[X(t + 1) − X(t) | X(t) = x, U (t) = u]
Fluid Model: x(t) = f (x(t), u(t))
˙
First-order Taylor series approximation:
Du h (x) = E[h(x + f (x, u, N ))] − h(x)
≈ f (x, u) · h (x)
12 / 26
29. Complex Networks
Approximate Dynamic Programming
ODE model from the MDP model, X(t + 1) − X(t) = f (X(t), U (t), N (t + 1))
Mean drift: f (x, u) = E[X(t + 1) − X(t) | X(t) = x, U (t) = u]
Fluid Model: x(t) = f (x(t), u(t))
˙
First-order Taylor series approximation:
Du h (x) = E[h(x + f (x, u, N ))] − h(x)
≈ f (x, u) · h (x)
A second-order Taylor series expansion
leads to a Diffusion Model.
12 / 26
30. Complex Networks
ADP for Stochastic Networks
Conclusions as of April 21, 2011
Stochastic Model: Q(t + 1) − Q(t) = B(t + 1)U (t) + A(t + 1)
d
Fluid Model: q(t) = Bu(t) + α Cost c(x, u) = |x|
dt
Relative value function h∗
Total cost value function J ∗
13 / 26
31. Complex Networks
ADP for Stochastic Networks
Conclusions as of April 21, 2011
Stochastic Model: Q(t + 1) − Q(t) = B(t + 1)U (t) + A(t + 1)
d
Fluid Model: q(t) = Bu(t) + α Cost c(x, u) = |x|
dt
Relative value function h∗
Total cost value function J ∗
Inventory model: Controlled work-release, controlled routing,
uncertain demand
q: Queue length evolves on R16 .
+
u: Scheduling/routing decisions —
demand 1
Convex relaxation
α: Mean exogenous arrivals of work
demand 2
B: Captures network topology
13 / 26
32. Complex Networks
ADP for Stochastic Networks
Conclusions as of April 21, 2011
Stochastic Model: Q(t + 1) − Q(t) = B(t + 1)U (t) + A(t + 1)
d
Fluid Model: q(t) = Bu(t) + α Cost c(x, u) = |x|
dt
Relative value function h∗
Total cost value function J ∗
Key conclusions – analytical
Stability of q implies stochastic stability of Q Dai, Dai & M. 1995
h∗ (x) ≈ J ∗ (x) for large |x| M. 1996–2011
In many cases, the translation of the optimal policy for q is
approximately optimal, with logarithmic regret M. 2005 & 2009
14 / 26
33. Complex Networks
ADP for Stochastic Networks
Conclusions as of April 21, 2011
Stochastic Model: Q(t + 1) − Q(t) = B(t + 1)U (t) + A(t + 1)
d
Fluid Model: q(t) = Bu(t) + α Cost c(x, u) = |x|
dt
Relative value function h∗
Total cost value function J ∗
Key conclusions – engineering
Stability of q implies stochastic stability of Q
Simple decentralized policies based on q Tassiulas, 1995 –
Workload relaxation for model reduction
M. 2003 –, following “heavy traffic” theory: Laws, Kelly, Harrison, Dai, ...
Intuition regarding structure of good policies
15 / 26
34. Complex Networks
ADP for Stochastic Networks
Workload Relaxations
R STO R∗
Inventory model: Controlled work-release, controlled routing, 50
uncertain demand
w2
demand 1
0
demand 2
-20
-20 0 50
w1
Workload process: W evolves on R2
Relaxation: Only lower bounds on rates are preserved
Effective cost: c(w) is the minimum of c(x), over all x consistent w.
¯
16 / 26
35. Complex Networks
ADP for Stochastic Networks
Workload Relaxations
R STO R∗
Inventory model: Controlled work-release, controlled routing, 50
uncertain demand
w2
demand 1
0
demand 2
-20
-20 0 50
w1
Workload process: W evolves on R2
Relaxation: Only lower bounds on rates are preserved
Effective cost: c(w) is the minimum of c(x), over all x consistent w.
¯
Optimal policy for fluid relaxation: Non-idling on region R∗
Optimal policy for stochastic relaxation: Introduce hedging
16 / 26
36. Complex Networks
ADP for Stochastic Networks
Policy translation
R STO R∗
Inventory model: Controlled work-release, controlled routing, 50
uncertain demand
w2
demand 1
0
demand 2
-20
-20 0 50
w1
Complete Policy Synthesis
1. Optimal control of relaxation
2. Translation to physical system:
2a. Achieve the approximation c(Q(t)) ≈ c(W (t))
¯
2b. Address boundary constraints ignored in fluid approximations
17 / 26
37. Complex Networks
ADP for Stochastic Networks
Policy translation
R STO R∗
Inventory model: Controlled work-release, controlled routing, 50
uncertain demand
w2
demand 1
0
demand 2
-20
-20 0 50
w1
Complete Policy Synthesis
1. Optimal control of relaxation
2. Translation to physical system:
2a. Achieve the approximation c(Q(t)) ≈ c(W (t))
¯
2b. Address boundary constraints ignored in fluid approximations
achieved using safety stocks.
17 / 26
38. Architectures for Adaptation & Learning
Singular Perturbations
Mean-Field Games Workload Relaxations
1
(individual state)
(ensemble state)
q1 q5
q2 q6
Agent 5 q 13 q 15
0 barely controllable q3 q7
Station 1
Station 2
d1
q8
Agent 4
q 16 q 14 q4 q9 q 12
-1 4 d2
0 1 2 3 4 5 6 7 8 9 10 x 10
Station 5
q 11 µ 10a q 10
µ 10b
Station 4 Station 3
Fluid model R STO R∗
50
w2
12.6
Di usion model
Average
Cost
12.4
Standard VIA 1
Initialized with quadratic Optimal policy 0.06
12.2
Initialized with optimal uid value function
12 0.05 0
11.8 0.04
11.6 0
0.03 -20
11.4
-20 0 50
0.02
11.2 w1
0.01
11
50 100 150 200 250 300 Iteration n
−1
−1 0 1
Adaptation & Learning
18 / 26
39. Architectures for Adaptation & Learning
Reinforcement Learning
Approximating a value function: Q-learning
ACOE Equation: min c(x, u) + Du h∗ (x) = η ∗
u
h∗ : Relative value function
η ∗ : Minimal average cost
19 / 26
40. Architectures for Adaptation & Learning
Reinforcement Learning
Approximating a value function: Q-learning
ACOE Equation: min c(x, u) + Du h∗ (x) = η ∗
u
h∗ : Relative value function
η ∗ : Minimal average cost
“Q-function”: Q∗ (x, u) = c(x, u) + Du h∗ (x)
Watkins 1989 ... “Machine Intelligence Lab”@ece.ufl.edu
19 / 26
41. Architectures for Adaptation & Learning
Reinforcement Learning
Approximating a value function: Q-learning
ACOE Equation: min c(x, u) + Du h∗ (x) = η ∗
u
h∗ : Relative value function
η ∗ : Minimal average cost
“Q-function”: Q∗ (x, u) = c(x, u) + Du h∗ (x)
Watkins 1989 ... “Machine Intelligence Lab”@ece.ufl.edu
Q-Learning: Given parameterized family {Qθ : θ ∈ Rd }.
Qθ is an approximation of the Q-function, or Hamiltonian Mehta & M. 2009
19 / 26
42. Architectures for Adaptation & Learning
Reinforcement Learning
Approximating a value function: Q-learning
ACOE Equation: min c(x, u) + Du h∗ (x) = η ∗
u
h∗ : Relative value function
η ∗ : Minimal average cost
“Q-function”: Q∗ (x, u) = c(x, u) + Du h∗ (x)
Watkins 1989 ... “Machine Intelligence Lab”@ece.ufl.edu
Q-Learning: Given parameterized family {Qθ : θ ∈ Rd }.
Qθ is an approximation of the Q-function, or Hamiltonian Mehta & M. 2009
Compute θ∗ based on observations — without using a system model.
19 / 26
43. Architectures for Adaptation & Learning
Reinforcement Learning
Approximating a value function: TD-learning
Value functions: For a given policy U (t) = φ(X(t)),
T
1
η = lim c(X(t), U (t)) dt
T →∞ T 0
Poisson’s equation: h is again called a relative value function,
c(x, u) + Du h (x) =η
u=φ(x)
20 / 26
44. Architectures for Adaptation & Learning
Reinforcement Learning
Approximating a value function: TD-learning
Value functions: For a given policy U (t) = φ(X(t)),
T
1
η = lim c(X(t), U (t)) dt
T →∞ T 0
Poisson’s equation: h is again called a relative value function,
c(x, u) + Du h (x) =η
u=φ(x)
TD-Learning: Given parameterized family {hθ : θ ∈ Rd }.
min{ h − hθ : θ ∈ Rd } Sutton 1988, Tsitsiklis & Van Roy, 1997
20 / 26
45. Architectures for Adaptation & Learning
Reinforcement Learning
Approximating a value function: TD-learning
Value functions: For a given policy U (t) = φ(X(t)),
T
1
η = lim c(X(t), U (t)) dt
T →∞ T 0
Poisson’s equation: h is again called a relative value function,
c(x, u) + Du h (x) =η
u=φ(x)
TD-Learning: Given parameterized family {hθ : θ ∈ Rd }.
min{ h − hθ : θ ∈ Rd } Sutton 1988, Tsitsiklis & Van Roy, 1997
Compute θ∗ based on observations — without using a system model.
20 / 26
46. Architectures for Adaptation & Learning
Reinforcement Learning
Approximating a value function: How do we choose a basis?
21 / 26
47. Architectures for Adaptation & Learning
Reinforcement Learning
Approximating a value function: How do we choose a basis?
Basis selection: hθ (x) = θi ψi (x)
ψ1 : Linearize
ψ2 : Fluid model with relaxation
ψ3 : Diffusion model with relaxation
ψ4 : Mean-field game
21 / 26
48. Architectures for Adaptation & Learning
Reinforcement Learning
Approximating a value function: How do we choose a basis?
Basis selection: hθ (x) = θi ψi (x)
ψ1 : Linearize
ψ2 : Fluid model with relaxation
ψ3 : Diffusion model with relaxation
ψ4 : Mean-field game
Examples: Decentralized control, nonlinear control, processor speed-scaling
1
1
Optimal policy 0.06 Approximate relative value function h
15 ∗
Fluid value function J
0.05
∗
Relative value function h
0.04
10
0
0
0.03
0.02
5
Agent 4 0.01
-1 −1
4 −1 0 1 0
0 5 10 x 10 0 5
Mean-Field Game Linearization Fluid Model
21 / 26
49. Next Steps
Nodal Power Prices in NZ: $/MWh
100
March 25:
50
0
4am 9am 2pm 7pm
Otahuhu
20,000
Stratford
March 26:
10,000
0 http://www.electricityinfo.co.nz/
4am 9am 2pm 7pm
Next Steps
22 / 26
51. Next Steps
Complex Systems
Mainly energy
Entropic Grid: Advances in systems theory...
Complex systems: Model reduction specialized to tomorrow’s grid
Short term operations and long-term planning
Resource allocation: Controlling supply, storage, and demand
Resource allocation with shared constraints.
Statistics and learning: For planning and forecasting
Both rare and common events
23 / 26
52. Next Steps
Complex Systems
Mainly energy
Entropic Grid: Advances in systems theory...
Complex systems: Model reduction specialized to tomorrow’s grid
Short term operations and long-term planning
Resource allocation: Controlling supply, storage, and demand
Resource allocation with shared constraints.
Statistics and learning: For planning and forecasting
Both rare and common events
Economics for an Entropic Grid: Incorporate dynamics and uncertainty
in a strategic setting.
How to create policies to protect participants on both sides of the
market, while creating incentives for R&D on renewable energy?
23 / 26
53. Next Steps
Complex Systems
Mainly energy
How to create policies to protect participants on both sides of the market,
while creating incentives for R&D on renewable energy?
Our community must consider long-term planning and policy, along with
traditional systems operations
24 / 26
54. Next Steps
Complex Systems
Mainly energy
How to create policies to protect participants on both sides of the market,
while creating incentives for R&D on renewable energy?
Our community must consider long-term planning and policy, along with
traditional systems operations
Planning and Policy, includes Markets & Competition
24 / 26
55. Next Steps
Complex Systems
Mainly energy
How to create policies to protect participants on both sides of the market,
while creating incentives for R&D on renewable energy?
Our community must consider long-term planning and policy, along with
traditional systems operations
Planning and Policy, includes Markets & Competition
Evolution?
24 / 26
56. Next Steps
Complex Systems
Mainly energy
How to create policies to protect participants on both sides of the market,
while creating incentives for R&D on renewable energy?
Our community must consider long-term planning and policy, along with
traditional systems operations
Planning and Policy, includes Markets & Competition
Evolution? Too slow!
24 / 26
57. Next Steps
Complex Systems
Mainly energy
How to create policies to protect participants on both sides of the market,
while creating incentives for R&D on renewable energy?
Our community must consider long-term planning and policy, along with
traditional systems operations
Planning and Policy, includes Markets & Competition
Evolution? Too slow!
What we need is Intelligent Design
24 / 26
58. Next Steps
Conclusions
The control community has created many techniques for understanding
complex systems, and a valuable philosophy for thinking about control
design
25 / 26
59. Next Steps
Conclusions
The control community has created many techniques for understanding
complex systems, and a valuable philosophy for thinking about control
design
In particular, stylized models can have great value:
Insight in formulation of control policies
Analysis of closed loop behavior, such as stability via ODE methods
Architectures for learning algorithms
Building bridges between OR, CS, and control disciplines
The ideas surveyed here arose from partnerships with researchers in
mathematics, economics, computer science, and operations research.
25 / 26
60. Next Steps
Conclusions
The control community has created many techniques for understanding
complex systems, and a valuable philosophy for thinking about control
design
In particular, stylized models can have great value:
Insight in formulation of control policies
Analysis of closed loop behavior, such as stability via ODE methods
Architectures for learning algorithms
Building bridges between OR, CS, and control disciplines
The ideas surveyed here arose from partnerships with researchers in
mathematics, economics, computer science, and operations research.
Besides the many technical open questions, my hope is to extend the
application of these ideas to long-range planning, especially in applications
to sustainable energy.
25 / 26
61. Next Steps
References
S. P. Meyn. Control Techniques for Complex Networks. Cambridge University Press,
Cambridge, 2007.
S. P. Meyn and R. L. Tweedie. Markov chains and stochastic stability. Second edition,
Cambridge University Press – Cambridge Mathematical Library, 2009.
S. Meyn. Stability and asymptotic optimality of generalized MaxWeight policies. SIAM J.
Control Optim., 47(6):3259–3294, 2009.
V. S. Borkar and S. P. Meyn. The ODE method for convergence of stochastic
approximation and reinforcement learning. SIAM J. Control Optim., 38(2):447–469, 2000.
S. P. Meyn. Sequencing and routing in multiclass queueing networks. Part II: Workload
relaxations. SIAM J. Control Optim., 42(1):178–217, 2003.
P. G. Mehta and S. P. Meyn. Q-learning and Pontryagin’s minimum principle. In Proc. of
the 48th IEEE Conf. on Dec. and Control, pp. 3598–3605, Dec. 2009.
W. Chen, D. Huang, A. A. Kulkarni, J. Unnikrishnan, Q. Zhu, P. Mehta, S. Meyn, and
A. Wierman. Approximate dynamic programming using fluid and diffusion approximations
with applications to power management. In Proc. of the 48th IEEE Conf. on Dec. and
Control, pp. 3575–3580, Dec. 2009.
26 / 26