on ai, markets and machine learning - computer scienceparkes/parkes.pdf · on ai, markets and...

On AI, Markets and Machine Learning

David C. Parkes

Computer Science

John A. Paulson School of Engineering and Applied Sciences

Harvard University

May 11, 2017

1 / 50

A Story about Design

Architecting the “outside” to make the “inside” easier, and promote

good outcomes.

Designing the rules of the game to bring AIs, and AIs and people,

together in useful ways.

Myriad applications: smart grid, participatory democracy, fair

division, resource allocation, matching, coalition formation, . . .

2 / 50

The Backdrop: ‘Value-aligned AI’

An AI-mediated society that strikes a balance between being fair,

utilitarian, representative, market-driven, . . .

Yikes! Clear that our field needs to be interdisciplinary (political

science, economic science, social psychology, ...)

And once we’ve understood all of this, we need to make our systems

scale.

3 / 50

scale.

3 / 50

scale.

3 / 50

To put a finer point on it...

There is concern about Silicon Valley “exporting its values to the

world,” with programs needing to encode values.

I think our community has a response, and it’s the coupling of:

Strong Agents:: Representing individual preferences, individual

values.

Normative System Design:: Aggregate inputs and make decisions in

a fair, representative, and utilitarian way.

4 / 50

To put a finer point on it...

There is concern about Silicon Valley “exporting its values to the

world,” with programs needing to encode values.

I think our community has a response, and it’s the coupling of:

Strong Agents:: Representing individual preferences, individual

values.

Normative System Design:: Aggregate inputs and make decisions in

a fair, representative, and utilitarian way.

4 / 50

Where I started...

5 / 50

Where I started...

6 / 50

Where I started...

Already autonomy, normative design, and connections w/ economics!

7 / 50

What I liked

Outside/inside: Design outside to promote simplicity inside, avoid

“wasteful counterspeculation.”

Precision: clearly stated design goals (norms), impossibility and

possibility results.

Apparent applicability, connection to practice.

8 / 50

(Aside) What I didn’t like

“Design an iterative combinatorial auction”

Regression testing by paper and pen.

Lucky break 1: Pat Harker suggested “go read Bertsekas’ work on

auction algorithms.”

Lucky break 2: Bikhchandani and Ostroy’98. An LP hierarchy, allowing

primal-dual formulations of CAs.

After a lot of sweat, iterative CAs as primal-dual algorithms.9 / 50

Fast forward a few years...

10 / 50

“I want to do research”

Rui Dong, a precocious freshman. It’s also my first week at Harvard.

We talk.

“But why is this computer science?”

To explain: let’s take the framework of mechanism design as an

example of the kinds of things that we do.

11 / 50

We talk.

11 / 50

We talk.

11 / 50

We talk.

11 / 50

Mechanism Design

n agents, set A of alternatives

Agent i’s value vi(a), for a ∈ A, with vi ∈ Vi.

Let V = V1 × . . .× Vn

A mechanism:

f : V 7→ A

Agents are self-interested. Properties of f studied in equilibrium:

g(v1, . . . , vn) = f(s∗1(v1), . . . , s∗n(vn))

12 / 50

Mechanism Design

Let V = V1 × . . .× Vn

A mechanism:

f : V 7→ A

g(v1, . . . , vn) = f(s∗1(v1), . . . , s∗n(vn))

12 / 50

Mechanism Design

Let V = V1 × . . .× Vn

A mechanism:

f : V 7→ A

g(v1, . . . , vn) = f(s∗1(v1), . . . , s∗n(vn))

12 / 50

The magical Revelation principle

An incentive compatible (IC) mechanism has a truthful equilibrium.

Simplifies!

The revelation principle: imposing IC is WLOG on mechanism design.

13 / 50

Simplifies!

14 / 50

Simplifies!

15 / 50

IC can be tricky to work with...

It’s a global property of a function.

Suppose v1 = w. Then for mechanism f to be IC, we need:

v1(f(w, v−1)) ≥ v1(f(x, v−1))

v1(f(w, v−1)) ≥ v1(f(y, v−1))

v1(f(w, v−1)) ≥ v1(f(z, v−1))

v1(f(w, v−1)) ≥ · · ·

Also need this for all agents, for all true valuations, and for all

valuations of others.

16 / 50

Another great idea: The VCG mechanism

Suppose agents other than 1 are truthful, let v1 denote report of 1.

Define f to solve

a∗ = arg maxa

(v1(a) +∑j 6=1

vj(a))

Charge agent 1 the amount −∑

j 6=1 vj(a∗). Utility:

v1(a∗) +

∑j 6=1

vj(a∗)

This provides IC! Can also modify payment, charge cost on others:∑j 6=1

vj(a−1)−∑j 6=1

vj(a∗)

17 / 50

Define f to solve

a∗ = arg maxa

(v1(a) +∑j 6=1

vj(a))

v1(a∗) +

∑j 6=1

vj(a∗)

vj(a−1)−∑j 6=1

vj(a∗)

17 / 50

Define f to solve

a∗ = arg maxa

(v1(a) +∑j 6=1

vj(a))

v1(a∗) +

∑j 6=1

vj(a∗)

vj(a−1)−∑j 6=1

vj(a∗)

17 / 50

It’s not really that strange a mechanism...

Equilibrium prices: [70,100] (balance supply and demand)

VCG charges cost on others: 70-0=70

This the min equil. price. Unaffected by winner’s bid.

18 / 50

The view from Economics

What can be done in principle?

19 / 50

Auctioning London Bus Routes

20 / 50

The View from Computer Science

How can we design real systems using ideas from mechanism design?

Need to scale-up: efficient preference elicitation, decentralized

computation, handle dynamics.

Lots more research— winner determination, design of bidding

languages, pricing algorithms, etc.

21 / 50

I: Efficient preference elicitation

Can we elicit just enough information to compute the ‘right

outcome’?

Modular architecture. Proxy agents query agents, continue until have

correct market outcome. (With S. Lahaie, F. Constantin)

22 / 50

I: Elicitation via Learning

monotone Bool functions

Representation class L

Target h : {0, 1}m 7→ RMembership query: h(x)

Equivalence query: h?

Learn poly(m, size(h))

free disposal

Bidding language L

Target vi : {0, 1}m 7→ RValue query: vi(x)

Demand query: (p, x)?

Solve poly(n,m, size(v1, vn))

23 / 50

free disposal

Bidding language L

23 / 50

free disposal

Bidding language L

Positive results for XOR, for polynomial language (substitutes).24 / 50

II: Removing the ‘Center’

Can we distribute the computation involved in f : V 7→ A?

What could possibly go wrong?

25 / 50

(w/ Jeff Shneidman, Boi Faltings, Adrian Petcu.)

Distribute computation of outcome rule and payments across agents

Need faithfulness: bring suggested strategy (= algorithm) into an ex

post Nash equilibrium

M-DPOP: adapted DPOP, a complete and optimal distr. opt alg to

make it faithful (via VCG ideas)

Crucial idea: The “partitioning principle”. Need ‘without i’ outcome

to be computed correctly whatever action of i.

26 / 50

III: Handling Dynamics

I remember teaching a class in around 2004. Someone asked about a

sequence of Vickrey auctions. Not IC!

Can we handle dynamic agent population, dynamically

changing information?

27 / 50

I remember teaching a class in around 2004. Someone asked about a

sequence of Vickrey auctions. Not IC!

Can we handle dynamic agent population, dynamically

changing information?

27 / 50

(w/ R. Cavallo, S. Singh, S. Seuken, R. Kleinberg, M. Mahdian, M. Hajiaghayi, F. Constantin)

With a monotone rule, can change the pricing rule. (Charge $40 in

this example.) Recover IC.

In a probabilistic model, follow optimal MDP policy, and use online

VCG mechanism, charge agent i

value− (V ∗(st)− V ∗(st,−i))

e.g., if v3 ∼ U(40, 60) then charge $50 to each of agents 1 and 2.28 / 50

(w/ R. Cavallo, S. Singh, S. Seuken, R. Kleinberg, M. Mahdian, M. Hajiaghayi, F. Constantin)

With a monotone rule, can change the pricing rule. (Charge $40 in

this example.) Recover IC.

In a probabilistic model, follow optimal MDP policy, and use online

VCG mechanism, charge agent i

value− (V ∗(st)− V ∗(st,−i))

e.g., if v3 ∼ U(40, 60) then charge $50 to each of agents 1 and 2.28 / 50

Broadening out: From Preferences to Actions

Thus far, living in this (abstract!) world

f : V 7→ A

But what if the inputs are not preferences but actions, for example:

Feedback on local places in a city

Pick-ups on a ride-sharing platform

Demand-response

A prototypical new design pattern is:

mechanism suggests x, sets payment schedule, agents decide whether to

do x or y, mechanism observes z, makes payment, ...

29 / 50

f : V 7→ A

Demand-response

29 / 50

f : V 7→ A

Demand-response

29 / 50

Eliciting Actions I: Google Local Guides

Question: how can we promote effort, elicit useful information?

30 / 50

Can we use ‘Peer prediction’ mechanisms?

Miller, Resnick and Zeckhauser, 2005

yes no( )agent 1

yes 0.5 0.1

no 0.1 0.3

yes no( )score

yes 1 0

no 0 2

Truthful reports maximize expected payment:

Ex2|x1[t1(x1, x2)] > Ex2|x1

[t1(x′1, x2)], for all x1, x′1

However: bad equilibria, need to know signal distribution.

(w/ J. Witkowski, D. Mandal, V. Shnayder, A. Agarwal, N. Shah,

R. Frongillo)

31 / 50

yes no( )agent 1

yes 0.5 0.1

no 0.1 0.3

yes no( )score

yes 1 0

no 0 2

Ex2|x1[t1(x1, x2)] > Ex2|x1

[t1(x′1, x2)], for all x1, x′1

R. Frongillo)

31 / 50

yes no( )agent 1

yes 0.5 0.1

no 0.1 0.3

yes no( )score

yes 1 0

no 0 2

Ex2|x1[t1(x1, x2)] > Ex2|x1

[t1(x′1, x2)], for all x1, x′1

R. Frongillo)

31 / 50

We’re not all alike (surprise!)

Zooniverse (exploding stars). E. Simpson et al. (2012)

Five distinct groupings of users. Labels -1 “not supernova.” +1

“possible supernova.” +3 “likely supernova.”

Group 1, clear in categorization of true 0s. Less sure about 1s.

Group 2 are “extremists.” Group 3 are “pessimists.” Group 4 are

“optimists.” Group 5 are “non-commitals.”

32 / 50

Robust Peer Prediction

(w/ A. Agarwal, D. Mandal and N. Shah.)

Multi-task

Learn the correlation structure, gives scoring scheme:

sgn(∆) =

+ + −+ + −− − +

; score =

Do this via clustering, and then estimate pairwise cluster

correlations. Incentive aligned!33 / 50

Multi-task

sgn(∆) =

+ + −+ + −− − +

; score =

Multi-task

sgn(∆) =

+ + −+ + −− − +

; score =

Eliciting Actions II: Ride-sharing platforms

34 / 50

Market Oriented Programming Strikes Back!

(with Fei Fang and Hongyao Ma)

Drivers and passengers are customers. Both ‘at will.’

Can only optimize by suggesting matches, prices.

Market-based optimization in practice (this time with real money!)

A challenge has been lack of smoothness of prices.

Can we design matching methods, and fair prices, so that

drivers always want to accept matched trips?

35 / 50

A view from Normative Design

Voluntary participation

Efficiency. Note externality. A trip from A to B has a side effect.

The car is now at B!

Fairness: any two drivers with the same (location,time) pair should

have same utility

In particular, we suggest the following interpretation of ‘smooth prices’:

Smoothness == equilibrium prices (in time and space)

We obtain anonymous, time-based trip prices. Come full circle, back to

primal/dual formulations...

36 / 50

A view from Normative Design

Voluntary participation

Efficiency. Note externality. A trip from A to B has a side effect.

The car is now at B!

Fairness: any two drivers with the same (location,time) pair should

have same utility

In particular, we suggest the following interpretation of ‘smooth prices’:

Smoothness == equilibrium prices (in time and space)

We obtain anonymous, time-based trip prices. Come full circle, back to

primal/dual formulations...

36 / 50

Towards Closing: History of Optimal Auctions

37 / 50

Can Deep Learning help?

Renaissance in neural networks (“deep learning”) for nonlinear

function approximation.

Robust tool chains (e.g., TensorFlow, GPUs).

Beginning to get attention in economics/EconCS (e.g., Hartford,

Wright, L-B’16a; Hartford, Lewis, L-B, Taddy’16b)

38 / 50

Automated (Data-driven) Mechanism Design

AMD first proposed by Conitzer and Sandholm, 2002.

Given: distribution v ∼ Fn

Learn an optimal auction (g, p)

Use neural networks for non-linear function approximation

(w/ H. Narasimhan, S. Agarwal, Z. Feng, P. Dutting.)

39 / 50

Deep learning for Automated mechanism design

Characterization-based architectures (MyersonNet, RochetNet), as well

as ‘agnostic networks’:

h(1)J1

h(R)JR

softmax

(a) Allocation network g

c(1)J ′1

c(T )1

c(T )2

c(T )J ′T

relu p1

relu p2

relu pn

(b) Payment network p

40 / 50

Example Allocation Rule (One buyer, multi-item)

(w/ Z. Feng, H. Narasimhan, P. Dutting.)

Recover existing optimal auctions.

Extend to new settings (e.g., 2 buyers, 2 items, continuous values.)

41 / 50

Looking ahead

We’re at a turning point for multi-agent systems. Many systems will

need to be designed.

The promise of agent-mediated electronic commerce is about to be

realized. Machina economica (w/ M. Wellman).

Crucial for value-aligned AI will be

strong agents, that can effectively represent individuals’

preferences and values

principled, normative design, that makes good tradeoffs

We will need to keep aware of other disciplines, collaborate.

42 / 50

Mentors

Lyle Ungar

Michael Wellman

Barbara Grosz

43 / 50

Friends

[thankyou!]

44 / 50

All of my Collaborators

[thankyou!]

45 / 50

Students and Postdocs!

[you’re the best!]

46 / 50

References

David C. Parkes and Lyle H. Ungar, “Iterative Combinatorial Auctions: Theory and Practice.” In Proc. 17th National

Conference on Artificial Intelligence (AAAI’00), 74-81 (2000)

Debasis Mishra and David C. Parkes, “Ascending Price Vickrey Auctions for General Valuations.” Journal of Economic

Theory, 132, 335-366 (2007)

Sebastien Lahaie and David C. Parkesm “Applying Learning Algorithms to Preference Elicitation.” In Proceedings of

the 5th ACM Conference on Electronic Commerce, 180-188 (2004)

Sebastien Lahaie, Florin Constantin, and David C. Parkes. “More on the Power of Demand Queries in Combinatorial

Auctions: Learning Atomic Languages and Handling Incentives.” In Proc. 19th International Joint Conference on

Artificial Intelligence (IJCAI’05), 959-964 (2005)

David C. Parkes and Jeffrey Shneidman “Distributed Implementations of Vickrey-Clarke-Groves Mechanisms.” In Proc.

3rd Int. Joint Conf. on Autonomous Agents and Multi Agent Systems, 261-268 (2004)

Jeffrey Shneidman and David C. Parkes, “Specification Faithfulness in Networks with Rational Nodes.” In Proc. 23rd

ACM Symp. on Principles of Distributed Computing (PODC’04), 88-97 (2004)

Adrian Petcu, Boi Faltings, and David C. Parkes, “MDPOP: Faithful Distributed Implementation of Efficient Social

Choice Problems” J. Artif. Intell. Res. 32: 705-755 (2008)

David C. Parkes and Satinder Singh, “An MDP-Based Approach to Online Mechanism Design.” In Proc. 17th Annual

Conf. on Neural Information Processing Systems (NIPS’03), 791-798 (2003)

47 / 50

David C. Parkes, Satinder Singh, and Dimah Yanovsky, “Approximately Efficient Online Mechanism Design.” In Proc.

18th Annual Conf. on Neural Information Processing Systems (NIPS’04), 1049-1056 (2004)

Ruggiero Cavallo, David C. Parkes, and Satinder Singh, “Optimal Coordinated Planning Amongst Self-Interested

Agents with Private State.” In Proceedings of the 22nd Conference on Uncertainty in Artificial Intelligence (UAI’06),

55-62 m(2006).

Ruggiero Cavallo and David C. Parkes, “Efficient Metadeliberation Auctions.” In Proc. 23rd AAAI Conference on

Artificial Intelligence (AAAI’08), 50-56 (2008)

Sven Seuken, Ruggiero Cavallo, and David C. Parkes, “Partially synchronized DEC-MDPs in Dynamic Mechanism

Design.” In Proc. 23rd AAAI Conference on Artificial Intelligence (AAAI’08), 162-169 (2008)

Ruggiero Cavallo, David C. Parkes, and Satinder Singh, “Efficient Mechanisms with Dynamic Populations and

Dynamic Types”. Working paper, 2010.

Mohammad T. Hajiaghayi, Robert Kleinberg, and David C. Parkes, “Adaptive Limited-Supply Online Auctions.” In

Proc. ACM Conf. on Electronic Commerce, 71-80 (2004).

Mohammad T. Hajiaghayi, Robert Kleinberg, Mohammad Mahdian, and David C. Parkes, “Online Auctions with

Re-usable Goods.” In Proc. 6th ACM Conf. on Electronic Commerce (EC’05), 165-174 (2005).

Florin Constantin and David C. Parkes, “Self-Correcting Sampling-Based Dynamic Multi-Unit Auctions.” In Proc. 10th

ACM Electronic Commerce Conference (EC’09), 89-98 (2009).

Jens Witkowski and David C. Parkes, “A Robust Bayesian Truth Serum for Small Populations.” In Proceedings of the

26th AAAI Conference on Artificial Intelligence (2012)

Jens Witkowski and David C. Parkes, “Peer Prediction without a Common Prior.” In Proceedings of the 13th ACM

Conference on Electronic Commerce (EC ’12), 964-981 (2012) 48 / 50

Jens Witkowski and David C. Parkes, “Learning the Prior in Minimal Peer Prediction.” In Proceedings of the 3rd

Workshop on Social Computing and User Generated Content (2013).

Victor Shnayder, Arpit Agarwal, Rafael Frongillo, and David C. Parkes, “Informed Truthfulness in Multi-Task Peer

Predictions.” In Proceedings of the 17th ACM Conf. on Economics and Computation (EC16), 179-196 (2016)

Arpit Agarwal, Debmalya Mandal, David C. Parkes, and Nisarg Shah, “Peer Prediction with Heterogeneous Users.” In

Proceedings of the 18th ACM Conference on Economics and Computation (2017)

Victor Shnayder, Rafael Frongillo, and David C. Parkes, “Measuring Performance Of Peer Prediction Mechanisms

Using Replicator Dynamics.” In Proceedings of the 25th International Joint Conference on Artificial Intelligence

(IJCAI’16), 2611-2617 (2016)

Hongyao Ma, Fei Fang and David C. Parkes, “Spatio-Temporal Pricing for Ride-Sharing Platforms”, Working paper

(2017)

Paul Dutting, Felix A. Fischer, Pichayut Jirapinyo, John Lai, Benjamin Lubin, and David C. Parkes, “Payment Rules

through Discriminant-Based Classifiers.” ACM Transactions on Economics and Computation, 3, 1:5 (2014)

Harikrishna Narasimhan and David C. Parkes, “A General Statistical Framework for Designing Strategy-proof

Assignment Mechanisms.” In Proceedings of the Conference on Uncertainty in Artificial Intelligence (2016)

Harikrishna Narasimhan, Shivani Agarwal, and David C. Parkes, “Automated Mechanism Design without Money via

Machine Learning.” In Proceedings of the 25th International Joint Conference on Artificial Intelligence (IJCAI’16),

433-439 (2016)

Paul Dutting, Zhe Feng, Harkrishna Narasimhan and David C. Parkes, “Optimal auctions through deep learning”,

arXiv:1706.03459 (2017)

49 / 50

Thank you

50 / 50

on ai, markets and machine learning - computer scienceparkes/parkes.pdf · on ai, markets and...

Documents