Title: PreBayesian Games
1Pre-Bayesian Games
- Moshe Tennenholtz
- TechnionIsrael Institute of Technology
2Acknowledgements
- Based on joint work with Itai Ashlagi, Ronen
Brafman and Dov Monderer.
3GT with CS flavor
- Program equilibrium /
- strong mediated equilibrium
- Ranking systems
- Non-cooperative computing
- Pre-Bayesian games
- Distributed Games
- Recommender systems for GT
-
-
4Modeling Uncertainty
- In game theory and economics the Bayesian
approach is mainly used. - Work in computer science frequently uses
non-probabilistic models. - Work on Pre-Bayesian games incorporates
game-theoretic reasoning into non-Bayesian
decision-making settings. -
5Pre-Bayesian Games
- Modeling and solution concepts in Pre-Bayesian
games. - Applications congestion games with incomplete
information. - Pre-Bayesian repeated/stochastic games as a
framework for multi-agent learning. -
6Games with Incomplete Information
7Model
8Model (cont.)
9Flexibility of the Model
10- Solution Concepts
- in Pre-Bayesian Games
11Dominant Strategies
12Ex-Post Equilibrium
13Safety-Level Equilibrium
-
- For every type play a strategy that maximizes
the worst-case payoff given the other players
strategies. - Worst case - over the set of possible states!
-
14Safety-Level Equilibrium
w
1-w
z
1-z
p
1-p
q
1-q
15Safety-Level Equilibrium (cont.)
16Other Non-Bayesian Solution Concepts
- Minimax-Regret equilibrium (Hyafil and Boutilier
2004) - Competitive-Ratio equilibrium
17Existence in Mixed Strategies
- Theorem
- Safety-level, Minimax-regret and
- Competitive-ratio equilibria exist in every
- concave pre-Bayesian Game.
- A concave pre-Bayesian game -
- - For every type the set of possible actions is
compact and convex (for every player) - - ui(?,) - concave function for every player i
- The proof follows by applying Kakutanis fixed
point theorem.
18Related Work On Non-Bayesian Solutions
- Safety-level equilibria
- Aghassi and Bertsimas (2004)
- Levin and Ozdenoren (2004)
- Pure safety-level equilbria
- Shoham and Tennenholtz (1992), Moses and
Tennenholtz (1992),Tennenholtz (1991) - Axiomatic Foundations
- Brafman and Tennenholtz (1996)
19Beyond Existence
20Modeling Congestion Settings
21Modeling Congestion Settings
- Examples
- Transportation engineering (Wardrop 1952, Beckman
et al. 1956) - Congestion games (Rosenthal 1973)
- Potential games (Monderer and Shapley 1996)
- Price of anarchy (Papadimitriou 1999, Tardos and
Roughgarden 2001) - Resource selection games with player specific
cost functions (Milchtaich 1996) - Local effect games (Leyton-Brown and Tennenholtz
2003) - .
-
-
22Where are we heading to?
- Our Goal Incorporate incomplete information to
congestion settings. - Type of uncertainties
- number of players
- job sizes
- network structure
- players cost functions
-
23Resource Selection Games with Unknown Number of
Players
24Resource Selection Games
25Symmetric Equilibrium
- Theorem
- Every resource selection game with
increasing resource cost functions has a unique
symmetric equilibrium - .
26Resource Selection Games with Unknown Number of
Players
27Uniqueness of Symmetric Safety Level Equilibrium
- - game with complete information
- - game with incomplete information
- Theorem
- Let ? be a resource selection system with
increasing resource cost functions.
has a unique symmetric safety-level
equilibrium. - The symmetric safety-level equilibrium profile is
. - is the unique symmetric equilibrium in
the game . -
28Is Ignorance Bad?
- K the real state, Kk , kltn
- Known number of players
- cost of every player
- Unknown number of players
- - cost of every player
-
-
- ?
29Is Ignorance Bad?
wj(k)wj(1)(k-1)dj
Main Theorem Let ? be a linear resource
selection system with increasing resource cost
functions. There exist an integer
such that for all 1.
2. All inequalities above are strict if
and only if there exists such
that
30Where is this Useful?
- Example
- Mechanism Design -
- Organizer knows the exact number of active
players. - Wishes to maximize social surplus -- will not
reveal the information.
31More Detailed Analysis
Theorem Let ? be a linear resource selection
system with increasing resource cost functions.
There exist an integer L such that for every
kgtL The minimal social cost in
attained with symmetric mixed-action profiles is
attained at . Consequently,
is minimized at n2k-1.
32Further Research
- Routing games with unknown number of players
- extension to general networks
- unique symmetric equilibrium exists in
a model where an - agent job can be split
- ignorance helps as long as nltk2
- Routing games with unknown job sizes
- extension to variable job sizes
- uncertainty about job sizes do not
change surplus in - several general settings
- ignorance helps where we have
uncertainty on both the - number of participants and
the job sizes -
-
-
- Minimax-regret equilibria in the different
congestion settings - Non-Bayesian equilibria in social choice settings
-
-
33Conclusions so far
- Non-Bayesian Equilibria exist in pre-Bayesian
Games. - Players are better off with common lack of
knowledge about the number of participants. - More generally, we show illuminating results
using non-Bayesian solution concepts in
pre-Bayesian games.
34- Non-Bayesian solutions for repeated (and
stochastic) games with incomplete information
efficient learning equilibrium.
35Learning in multi-agent systems
- Multi-Agent Learning lies in the intersection of
Machine Learning/Artificial Intelligence and Game
Theory - Basic settings
- A repeated game where the game (payoff
functions) is initially unknown, but may be
learned based on observed history. - A stochastic game where both the stage games
and the transition probabilities are initially
unknown. - What can be observed following an action is part
of the problem specification. - No Bayesian assumptions!
36The objective of learning in multi-agent systems
- Descriptive objective how do people behave/adapt
their behavior in (e.g. repeated) games? - Normative objective can we provide the agents
with advice about how they should behave, to be
followed by rational agents, which will also
lead to some good social outcome?
37Learning in games an existing perspective
- Most work on learning in games (in machine
learning/AI extending upon work in game theory),
deals with the search for learning algorithms
that if adopted by all agents will lead to
equilibrium. - (another approach regret minimization will be
discussed and compared to later).
38Re-Considering Learning in Games
- But, why should the agents adopt these learning
algorithms? - This seems contradicting to the whole idea of
self-motivated agents (which led to considering
equilibrium concepts). -
39Re-Considering Learning in Games
-
- (New) Normative answer The learning algorithms
themselves should be in equilibrium! - We call this form of equilibrium Learning
Equilibrium, and in particular we consider
Efficient Learning Equilibrium (ELE). - Remark In this talk we refer to optimal ELE
(extending upon the basic ELE we introduced) but
use the term ELE.
40Efficient Learning EquilibriumInformal
Definition
- The learning algorithms themselves are in
equilibrium. It is irrational for an agent to
deviate from its algorithm assuming that the
others stick to their algorithms, regardless of
the nature of the (actual) game that is being
played. - If the agents follow the provided learning
algorithms then they will obtain a value that is
close to the value obtained in an optimal (or
Pareto-optimal) Nash equilibrium (of the actual
game) after polynomially many iterations. - It is irrational to deviate from the learning
algorithm. Moreover, the irrationality of
deviation is manifested within a polynomial
number of iterations.
41- Efficient Learning Equilibrium is a form of
ex-post equilibrium in Pre-Bayesian repeated games
42Basic Definitions
- Game GltN1,,n,S1,.,Sn,U1,.,Ungt
- UiS1 ? ? Sn? R - utility function
for i - ?(Si) mixed strategies for i.
- A tuple of (mixed) strategies t(t1,,tn) is a
Nash equilibrium if - ?i ?N, Ui(t) ? Ui(t1,,ti-1,t,ti1,,tn)
for every t ? Si - Optimal Nash equilibrium maximizes social
surplus - (sum of agents
payoffs) - val(t,i,g) the minimal expected payoff that
may be obtained by i - when employing t in the game g.
- A strategy t ? ?(Si) for which val(.,i,g) is
maximized is a - safety- level strategy (or, probabilistic
maximin strategy ), and its value is the
safety-level value.
43Basic Definitions
- R(G) -- repeated game with respect to a
(one-shot) game G. - History of player i after t iterations of R(G)
- Perfect monitoring Hti ((a1j, ,
anj),(p1j,,pnj))tj1 --- - a player can observe all previously chosen
actions and payoffs - Imperfect monitoring Hti ((a1j, ,
anj),pij)tj1 --- - a player can observe previously chosen
actions (of all players) - and payoffs of i.
- Strictly imperfect monitoring Hti
(aij ,pij)tj1 --- - a player can observe only its own
payoffs and actions. - Possible histories for agent i Hi??t1
Hti ? ? - Policy for agent i ?Hi? ?(Si)
- Remark in the game theory literature the term
perfect monitoring is used to refer to the
concept of imperfect monitoring above -
44Basic Definitions
- Let G be a (one-shot) game, let MR(G) be the
corresponding repeated game, and let n(G) be an
optimal Nash-equilibrium of G. Denote the
expected payoff of agent i in that equilibrium by
NVi(n(G)). -
- Given M R(G) and a natural number T, we denote
the expected T-step undiscounted average reward
of player i when the players follow the policy
profile (?1 ,,?i,,?n) by Ui(M,?1
,,?i,,?n,T). -
- Ui(M,?1 ,,?i,,?n)liminfT ? ?Ui(M,?1
,,?i,,?n,T)
45Definition (Optimal) ELE (in 2-person repeated
game)
(?,?) is an efficient learning equilibrium with
respect to the class of games ? (where each
one-shot game has k actions) if for every ? gt 0,
0 lt ? lt1, there exists some Tgt0, where T is
polynomial in 1/ ?, 1/ ?, and k, such that with
probability of at least 1- ? (1) If player
1(resp. 2) deviates from ? to ? (resp. from ? to
?) in iteration l, then U1(M,(?,?) ,lt)
?U1(M,(?,?) ,lt) ? (resp. U2(M,(?,?) ,lt)
?U2(M,(?,?) ,lt) ?) for every t ? T and for
every repeated game MR(G) ? ? . (2) For every t
? T and for every repeated game MR(G) ? ? ,
U1(M,(?,?) ,t) U2(M,(?,?) ,t) ?
NV1(n(G))NV2(n(G)) - ? for an optimal (surplus
maximizing) Nash equilibrium n(G).
46The Existence of ELE
Theorem Let M be a class of repeated games.
Then, there exists an ELE w.r.t. M given perfect
monitoring. The proof of the above is
constructive and use ideas of our Rmax algorithm
(the first near-optimal polynomial time algorithm
for reinforcement learning in stochastic
games)the folk-theorem in economics.
47The ELE algorithm
- For ease of presentation assume that the payoff
functions are non-negative and are bounded by
Rmax. - Player 1 performs action ai one time after the
other for k times, for all i1,2,...,k. - In parallel, player 2 performs the sequence of
actions (a1,,ak) k times. - If both players behaved according to the above
then an optimal Nash equilibrium of the
corresponding (revealed) game is computed, and
the players behave according to the corresponding
strategies from that point on. If several such
Nash equilibria exist, one is selected based on a
pre-determined arrangement. - If one of the players deviated from the above,
we shall call this player the adversary and the
other player the agent, and do the following - Let G be the Rmax-sum game in which the
adversary's payoff is identical to his payoff in
the original game, and where the agent's payoff
is Rmax minus the adversary payoffs. Let M
denote the corresponding repeated game. Thus, G
is a constant-sum game where the agent's goal is
to minimize the adversary's payoff. Notice that
some of these payoffs will be unknown (because
the adversary did not cooperate in the
exploration phase). The agent now plays according
to the following
48The ELE algorithm (cont.)
- Initialize Construct the following model M' of
the repeated game M, where the game G is replaced
by a game G' where all the entries in the game
matrix are assigned the rewards (Rmax,0) (we
assume w.l.o.g positive payoffs, and also assume
the maximal possible reward Rmax is known). - We associate a boolean valued variable with each
joint-action assumed,known. This variable is
initialized to the value assumed. - Repeat
- Compute and Act Compute the optimal
probabilistic maximin of G' and - execute it.
- Observe and update Following each joint action
do as follows - Let a be the action the agent performed and
let a be the adversary's action. - If (a,a') is performed for the first
time, update the reward associated with - (a,a') in G', as observed, and mark it
known.
49Imperfect Monitoring
Theorem There exist classes of games for which
an ELE does not exist given imperfect
monitoring. The proof is based on showing that
you can not get the values obtained in the Nash
equilibria of the following games, when you dont
know initially what game you play, and can not
observe the other agents payoff
50The Existence of ELE for Imperfect Monitoring
Settings
Theorem Let M be a class of repeated symmetric
games. Then, there exists an ELE w.r.t. M given
imperfect monitoring.
51The Existence of ELE for Imperfect Monitoring
Settings Proof Idea
Agents are instructed to explore the game matrix.
If it has been done without deviations,
action profiles (s,t) and (t,s) with optimal
surplus are selected to be played indefinitely
when (s,t) is played on odd iterations and (t,s)
is played on even iterations. If there has
been a deviation then we remain with the problem
of effective and efficient punishment. Notice
that here an agent does not learn another agents
payoff in an entry once it is played!
52The Existence of ELE for Imperfect Monitoring
Settings Proof Idea (cont.)
Assume the row agent is about to punish the
column agent. We say that a column associated
with action s is known if the row agent knows its
payoff for any pair (t,s). Notice that at each
point the squared sub-matrix which corresponds to
actions associated with known columns has the
property that the row agent knows all payoffs of
both agents in it. With some small probability
the row agent plays a random action, and
otherwise plays the probabilistic maximin
associated with the above (known) squared
sub-matrix where its payoffs are the complement
to 0 of the column agent payoffs. Many missing
details and computations.
53Extensions
- The results are extended to n-person games and
stochastic games, providing a general solution to
the normative problem of multi-agent learning.
54ELE and Efficiency
Our results for symmetric games imply that we can
get the optimal social surplus as a result of the
learning process, where the learning algorithms
are in equilibrium! This is impossible in
general games without having side payments as
part of the policies, which leads to another
version of the ELE concept.
55Pareto ELE
Given a 2-person game G, a pair (a,b) of
strategies is (economically) efficient, if
U1(a,b)U2(a,b)maxs?S1,t?S2 (U1(s,t)U2(s,t)) Obt
aining economically efficient outcomes is in
general impossible without side payments (the
probabilistic maximin value for i may be higher
than what he gets in the economically efficient
tuple). Side payments an agent may be asked
to pay the other as part of its policy. If its
payoff at particular point is pi, and the agent
pays ci then the actual payoff/utility is pi-ci.
Pareto ELE is defined similarly to (Nash) ELE
with the following distinctions 1. The
agents should obtain an average total reward
close to the sum of their rewards in an
efficient outcome. 2. Side payments are
allowed as part of the agents policy.
56Pareto ELE
Theorem Let M be a class of repeated games.
Then, there exists a Pareto ELE w.r.t. M given
perfect monitoring. Theorem There exist classes
of games for which a Pareto ELE does not exist
given imperfect monitoring.
57Common Interest Games
A game is called a common-interest game if for
every joint-action all agents receive the same
reward. Theorem Let Mc be the class of
common-interest repeated games in which the
number of actions each agent has is a. There
exists an ELE for Mc under strict imperfect
monitoring. The above result is obtained for the
general case where there are no a-priori
conventions on agents ordering or strategies
ordering.
58Efficient Learning Equilibrium and Regret
Minimization
- The literature on regret minimization attempts to
find a best response for arbitrary action
sequences of an opponent. - Notice that in general an agent can not devise
best response against an adversary whose action
selection depends on the agents previous
actions. - In such situations it is hard to avoid
equilibrium concepts. - Efficient Learning Equilibrium requires that
deviations will be irrational considering any
game from a given set of games, and therefore has
the flavor of ex-post equilibrium.
59Stochastic Game
0.3
a2adver.
a1adver.
0.5
a1agent
1
a2agent
0.5
0.7
0.4
0.6
60SGs Are An Expressive Model
- SGs are more general than Markov decision
processes and repeated games - Markov decision process the adversary has a
single action - Repeated games a unique stage game
61Extending ELE to stochastic games
Let M be a stochastic game and let ? gt 0, 0 lt ?
lt1. Let vi(M,?) be the ?-return mixing time of a
probabilistic maximin (safety level) strategy for
agent i. Consider a stochastic game, Mi, which
is identical to M except that the payoffs of
player i are taken as the complement to Rmax of
the other player's payoff. Let vi'(Mi, ?) be the
?-return mixing time of an optimal policy
(safety-level strategy) of i in that game.
Consider also the game M, where M is a Markov
decision process, which is isomorphic to M, but
where the (single) player's reward for the action
(a,b) in state s is the sum of the players'
rewards in M. Let Opt(M') be the value of an
optimal policy in M'. Let vc(M, ?) be the
?-return mixing time of that optimal policy (in
M). Let v(M, ?)max(v1(M, ?),v1(M, ?) , v2(M,
?),v2(M, ?),vc(M, ?))
62Extending ELE to stochastic games
A policy profile (?,?) is a Pareto efficient
learning equilibrium w.r.t. the class M of
stochastic games if for every ? gt 0, 0 lt ? lt1,
and M ? ?, there exists some Tgt0, where T
is polynomial in 1/?, 1/ ?, the size of M, and
v(M, ?), such that with probability of at least
1- ? (1) for every t ? T, U1(M,?,?,t)
U2(M,?,?,t) ? (1- ?)(Opt(M')) - ? for i1,2 (2)
if player 1 (resp. 2) deviates from ? to ?(resp.
from ? to ?) in iteration l, then
U1(M,?,?,lt)? U1(M,?,?,lt) ? (resp.
U2(M,?,?,lt)? U2(M,?,?,lt) ?) ) Theorem
Given a perfect monitoring setting for stochastic
games, there always exists a Pareto ELE.
63The R-max Algorithm
- R-max is the first near-optimal efficient
reinforcement learning algorithm for stochastic
games. In particular, it is applicable to
(efficiently) obtaining the safety-level value in
stochastic games where the stage games and
transition probabilities are initially unknown - Therefore, when adopted by all agents, R-max
determines an ELE in zero-sum stochastic games. - Efficiency is measured as a function of the
mixing time of the optimal policy in the known
model.
64The R-max Algorithm
- A model-based learning algorithm utilizing an
optimistic, fictitious model - Model initialization
- States original states 1 fictitious state
- All game-matrix entries are marked unknown
- All joint actions lead to the fictitious state
with probability 1 - The agents payoff is Rmax everywhere (the
adversarys payoff plays no role, 0 is fine)
65Initial Model
Fictitious Stage Game
- Unknown
1
Real Stage Games
66The Algorithm (cont.)
- Repeat
- Compute optimal policy
- Execute current policy
- Update model
67Model Update
- Occurs after we play a joint action corresponding
to an unknown entry - Record payoff in matrix (once only)
- Record the observed transition
- Once enough transitions from this entry are
recorded - Update the transition model based on the observed
frequencies - Mark the entry as known
- Recompute the policy
68The Algorithm (cont.)
- Repeat
- Compute optimal T-step policy
- Execute current policy
- Update model an entry is known when it has
been visited - times.
69Main Theorem
- Let M be an SG with N states and k actions. Let
egt0 and 0ltdlt1 be constants denoting desired error
bounds. Denote the policies for M whose e-return
mixing time is T by pM(e,T), and the optimal
expected return achievable by such policies by
OptM(p(e,T)) (i.e., the best value of a policy
that e-mixes in time T).
70Main Theorem (cont.)
- Then, with probability no less than 1-d, the
R-max algorithm will attain an actual average
return of no less than OptM(p(e,T))-e within a
number of steps polynomial in - N,T,k,1/d,1/e.
71Main Technical Contribution Implicit Explore or
Exploit (IEE)
- R-max either explores efficiently or exploits
efficiently - The adversary can influence whether we exploit
efficiently or explore - But, it cannot prevent us from doing one of the
two
72Conclusion (ELE)
- ELE captures the requirement that the learning
algorithms themselves should be in equilibrium. - Somewhat surprisingly, (optimal) ELE exists for
large classes of games. The proofs are
constructive. - ELE can be viewed as ex-post equilibrium in
repeated pre-Bayesian games with (initial) strict
uncertainty about payoffs. - The results can be extended to stochastic games
(more complicated, and need to refer to mixing
time of policies in the definition of
efficiency).
73Conclusion
- Pre-Bayesian Games are a natural setting for the
study of multi-agent interactions with incomplete
information, where there is no exact
probabilistic information about the environment. - Natural solution concepts such as ex-post
equilibrium can be extended to non-Bayesian
equilibrium (such as safety-level equilibrium)
which always exist. - The study of non-Bayesian equilibrium leads to
illuminating results in areas connecting CS and
GT.
74Conclusion (cont.)
- There are tight connection between Pre-Bayesian
repeated games and multi-agent learning. - Equilibrium of learning algorithms can be shown
to exist in rich settings. ELE is a notion of
ex-post equilibrium in Pre-Bayesian repeated
games. - The study of Pre-Bayesian games is a rich,
attractive, and illuminating research direction!
75Our research agenda GT with CS flavor
- Program equilibrium
- Ranking systems
- Non-cooperative computing
- Pre-Bayesian games
- Distributed Games
- Recommender systems for GT
-
-
76GT with CS flavor re-visiting equilibrium
analysis
- Program equilibrium
- CS brings the idea that strategies can be
of low capability (resource bounds), but also of
high capability programs can serve both as data
and as a set of instructions. This enables to
obtain phenomena observed in repeated games in
the context of one-shot games.
77GT with CS flavor re-visiting social choice
- Ranking systems
- The Internet suggests the need to
extend the theory of social choice to the context
where the set of players and the set of
alternatives coincide and transitive effects are
taken into account. This allows to treat the
foundations of page ranking systems and of
reputation systems (e.g. an axiomatization of
Googles PageRank).
78GT with CS flavor re-visiting mechanism design
- Non-cooperative computing
- Informational mechanism design where
goals are informational states, and agents
payoffs are determined by informational states is
essential in order to deal with distributed
computing with selfish participants. This allows
to answer the question of which functions can be
jointly computed by self-motivated participants.
-
79GT with CS flavor action prediction in one-shot
games
- Recommender systems for GT
- Find correlations between agents behaviors
in different games, in order to try and predict
an agents behavior in a game (he has not played
yet) based on his behavior in other games. This
is a useful technique when e.g. selling books in
Amazon, and here it is suggested for action
prediction in games, with surprisingly great
initial success (an experimental study). -
80GT with CS flavor incorporating distributed
systems features into game theoretic models
-
- Distributed Games
-
- The effects of asynchronous
interactions - The effects of message syntax and the
- communication structure on
implementation - The effects of failures.
-
81GT with CS flavor revisiting uncertainty and in
games and learning
- Pre-Bayesian games
- This talk.