Robust Mechanisms for Information Elicitation Aviv Zohar & Jeffrey S. Rosenschein Overview of the talk • • • • • • Motivation – how to pay for information Scoring Rules Mechanisms for information elicitation Robust mechanisms for a single agent Multi-agent extensions Conclusion Motivation - How to Pay for Information • Alice wishes to know the weather in TelAviv – This cannot be predicted in advance by anyone! • She’s only interested in the chance of rain. • There are two options: Paying for Information • Bob lives in Tel-Aviv. • He can go outside and check the weather. • Getting the information costs him some effort. A cost of c. • He wants Alice to pay him for his efforts. Paying for Information • If Alice pays him c$ no matter what he says he can just make something up. • If Alice pays him c1$ for saying rain, and c2$ for saying clear weather, he will pick the larger payment every time. • Conclusion: Alice has to have some way to verify the information. • Example: The weather in Jerusalem is correlated with the weather in Tel-Aviv • In the real world we often buy unverified information. We are usually playing a repeated game. How to Pay for Information • Bob knows that Tel-Aviv is usually sunny. • His beliefs affect the cost-benefit analysis. • Alice needs to take Bob’s beliefs into consideration when deciding on payments. • Does she know what Bob believes? Usually only approximately! • Can she find a payment scheme to Bob that will be robust against small changes in belief? Information Elicitation vs. Preference Elicitation. • Information Elicitation: – – – – Knowledge is changing hands. The seller only cares about payment. Not interested in how the knowledge is to be used. The buyer wants the truth! • Preference Elicitation: – Information is just the means to an end: Achieving some optimal outcome. – The outcome is the bottom line. – The mechanism has more freedom of action – can control outcome as well as payments. (EXAMPLE: Auctions) The Direct Revelation Principle • If any mechanism exists for the problem then there is a mechanism in which the participants reveal everything. Outcome+ Payments Outcome+ Payments A1 Mech. A1 a1 Mech. A3 a2 A2 A2 a3 A3 Direct Revelation • Direct revelation can allow us to get over the problem of learning bob’s beliefs. • We can ask Bob to reveal everything: – The information to be sold – His beliefs about probabilities. • But… – We don’t want to reveal everything. – Information is what Bob sells! – No trusted third party. Proper Scoring Rules • A way to evaluate a probabilistic prediction. • For a prediction p, and a final outcome o we shall pay: S(p,o). • A proper scoring rule is one in which telling a more accurate prediction gives a higer payment: • Eo~p[S(p,o)] > Eo~p[S(q,o)] Scoring Rules • Eo~p[S(p,o)] > Eo~p[S(q,o)] • There are lots of functions S(.,.) that fulfill this condition. • Example: Logarithmic payments: S(p,i) = log(pi) In which case: Ei~ p [S(p, i)] pi log( pi ) pi E i ~ p [S(p, i)] - E i ~ p [S(q, i)] pi log 0 qi Scoring Rules • Predictions are given in the form of probability distributions. • How do we combine the predictions of two different experts that have access to different sources of information? • We need a model of how their information interacts. Our Model • Seller i owns a random variable Xi that it pays ci to discover. • Buyer owns a random variable Ω • After it learns about values x’ from the sellers and a value ω, it pays the sellers uiω,x’ • The variables X1,X2,…,Ω are presumably not independent • We assume that they are governed by probability distribution px1,x2,…,ω • Now we know how to combine information from several sources. Pr(ω|X1,X2…) The Model 1 X1 c1 x1 X2 u , x '1 , x '2 Seller 1 u2 , x '1 , x '2 c2 x2 x'1 Seller 2 x' 2 Buyer Ω The Requirements from a Proper Mechanism (Single Agent) 1. Truth-telling: The truth is more profitable than any lie. x x' p ,x u , x p , x u , x ' 2. Investment: Knowing is better than guessing. x' p ,x u , x c p , x u , x ' ,x ,x 3. Individual Rationality: There is a positive expected gain from participating. p ,x ,x u , x c Finding a Mechanism • We assume P is known. • The constraints are all linear in the payments u. • We can find a payment scheme using some LP solver. • We can optimize the cost too: min p ,x u , x ,x • When can we find a good mechanism? • What is the optimal cost? The Truth is Enough • Suppose we have some set of payments that satisfies the truthfulness constraints: x x' p ,x (u , x u , x ' ) 0 • We can scale and shift it u , x u , x To satisfy the other constraints. x' p ,x (u , x u , x ' ) c ,x u , x c ,x p ,x The Truthfulness Constraints • Let’s define: u x (u1, x ,..., un , x ) p x ( p1, x ,..., pn , x ) vx, x ' u x u x ' • Now we get: x x' • And also: p ,x (u , x u , x ' ) 0 px vx , x ' 0 px ' vx, x' 0 • We need every pair of vectors px,px’ to be linearly separated by vx,x’ A Geometric Interpretation of Truthfulness ω2 px px vx, x ' 0 Vx,x’ px ' vx, x' 0 • Notice that there can be many ways to select the separating plane px’ ω1 Existence of a Mechanism • A mechanism exists if and only if all vectors px are pair-wise independent. – One direction is easy: we can’t separate vectors that are linearly dependent. – For the other direction: show a working mechanism: px – Setting u x does the trick. px p x (u x u x ' ) p x px px' px px' p x p x px' 0 px' Robust Mechanisms • We return to the case where P is not known exactly. px pˆ x x • We assume ε is small (according to L∞). • certain solutions may be better than others ω2 px px’ ω1 Robustness of a Specific Payment Scheme • A conservative definition: A payment scheme u will be considered ε-robust with regard to distribution if it is p̂ proper for every distribution p pˆ for which • How do we find the robustness level of a payment scheme? – Find the minimal ε for which a constraint is violated. Robustness of a Payment Scheme • The robustness of one of the truthfulness constraints can be found by solving: min ( pˆ , x , x ) (u , x u , x' ) 0 ,x 0 ,x pˆ , x , x 0 ,x • After solving a similar program for every constraint, take the smallest ε found. Finding a Robust solution • Given an ε, all ε-robust solutions form a convex set. • This is a stochastic programming problem. – Find a solution to a mathematical program with uncertainty regarding the constraints. • The ellipsoid algorithm needs only a separation oracle in order to optimize over the set of solutions. • A separation oracle provides a linear separator between any non-solution and the set of solutions. • We’ve just seen how to find one! Find a constraint that the payments + a perturbation violate. The full stochastic program: min p ,x u , x ,x p x x' x' p ,x ,x u , x p , x u , x ' u , x c p , x u , x ' ,x p ,x u , x c ,x 0 ,x ,x p , x pˆ , x , x 0 ,x ,x Robust Mechanisms • Definition: The robustness level of a problem p is the largest ε that can be set as the robustness of a solution. • How can we find it? • Use binary search. – The robustness level is somewhere between 0 and 1. – Test at any wanted ε in between by trying to actually find an ε-robust solution. – Then, update the boundaries according to the answer. A Bound for Problem Robustness • Problem robustness is only is only bounded by the truthfulness constraints. – Again, shifting and rescaling takes care of the other constraints. • A simple bound can be derived: min pˆ x pˆ x x, x ' x, x ' tr * x, x ' x, x' pˆ x pˆ x ' pˆ x pˆ x ' 2 ω2 px εx ε x’ px’ Mechanisms for Multiple Sellers • Collusion between agents is a critical matter. • If they can move payments and share information, we can treat them as one agent with multiple sources of information. • An exponential number of constraints is needed. • Tension within the group may limit their collusion. • From here on we assume no collusion is possible. Mechanisms for Multiple Sellers Two main options: 1. Mechanisms that work in only in equilibrium. – Truth telling is profitable only when everyone else does it. – Payments are conditioned on all the information – Other equilibriums may exist. 2. Dominant strategy mechanisms. – It is always better to tell the truth. – Payments are conditioned on the agent’s own information only (And the verifier). A Simple Example (2x2x2 ) x1 x2 Pr(x1,x2) Pr(ω=1|x1,x2) 0 0 1 1 0 1 0 1 1/4 1/4 1/4 1/4 0 1- δ 1 δ Pr(ω|x2=1)=Pr(ω|x2=0) • Hence no dominant strategy mechanism for player 2 • But a mechanism in equilibrium exists. A Simple Example (2x2x2 ) x1 x2 Pr(x1,x2) Pr(ω=1|x1,x2) 0 0 1 1 0 1 0 1 1/4 1/4 1/4 1/4 0 1- δ 1 δ • The variable Ω is slightly biased to agree with player 1’s variable. • We can have a dominant strategy mechanism for player 1. Robust Mechanisms for Many Sellers • Dominant strategy mechanism – Just like the single agent case. May not always exist. • Mechanisms that work in equilibrium- problematic. • An equilibrium is a best response to a best response. • A player must believe that its counterpart will play the equilibrium strategy. • This only happens if it believes that the other believes that it will play the equilibrium. • And so on… Belief Hierarchies • Assume player A believes the probability is p • player B might conceivably believe it’s p’ • Furthermore it may believe that A believes it is p’’. • p’’ may be far from p, and we get further away with every step. P’ P P’’ What can we do? • We can consider bounded players. Only look some distance into the belief hierarchy. • We can create finite belief hierarchies. – The first player has a dominant strategy. – The payment to second player depends only on the first. – Payment to the third only on the previous two – Etc. • Every player considers just beliefs of players that precede him. • They do not care about his beliefs. No loops. Finite belief hierarchies • Only a single equilibrium. • Very reasonable that it will be played. • Such mechanisms might not always exist • The extreme case: – All agents have a dominant strategy mechanism. Conclusion • Designing information elicitation mechanisms: – Easy for one agent. – Can be extended easily to robust mechanism – Complicated for many agent. – Robust extension is unclear in equilibrium. – Collusion makes the design even harder.