Incentive Centered Design : Hidden Action
Some results from class:
1. When principal (P) is neutral and agent A is risk averse - fixed payment contracts.
and the principal bears all the uncertainty
2. When the P is risk averse and the agent A is risk neutral then : "Sell the project to the agent". i.e. franchise contarcts.
e.g. the agent pays McDonalds to own the franchise - McD gets a fixed payment and the franchise takes all the risk.
However in real life both principal and agent will have some kind of risk aversion and so in real life the contract will be something between the two.
The contracting timeline:
1. principal designs contract
2. agent accepts of rejects contract (IR contract - forms one variable in the lagrangian later)
3. how hard is agent going to work (Incentive Compatibility Constraint) (forms another variable for the lagrangian later)
4. Uncertainty is resolved, i.e. we have some outcome; important point to consider
how do payoffs depend on the outcome of the project - design decision.
5. contract payoffs are made - and happily ever after :)
The timeline allows us to use backward recursion to design optimal contracts.
Lots of weird, torturous equations later :)
IR > 0 => how costly is it to get the agent to just sign ujp
IC > 0 => how costly is the lack of monitoring of information for the principal
and most importantly
t(x)= optimal payment to agent
t(x) depends directly on the likelihood ratio : i.e. the ratio p(i,L)/p(i,h)
{ note here:
derivative of utility of agent t(u(x)) is decreasing
=> as t(x) increases => du decreases => 1/du increases
}
where p(i,h) represents how likely it is to observe result i (state i of the world) if agent worked hard (h)
so if p(i,h)/p(i,l) is small we are much more likely to see results x(i) if effort to high i.e. e=e(h)
the takeaway is since we can't measure effort we measure some quantity that encapsulates effort. here x. so effective you are being paid not for e but for the information encapsulated and communicated by x.
Thus we want to motivate profitable behavior for principal my making working efficiently attractive for agent well - and be able to monitor that in some observable outcome.
For example Varszegi paid his employees at fotex four times the market price in the new hungarian regime - their incentive ? - to keep their jobs by showing Varszegi that they were valuable to his business by working hard. So the profitability of the business encapsulated some information about the work of the employees and Varszegi made a killing. Stock options in companies have the same purpose.
The design question for me is: how many ways can we leverage this in interaction design. How can the interface encapsulate knowledge about user actions and then reflect that to either motivate others or motivate the same user; Are recommender systems at least in part an instantiation of this - or not ? i'll have to wait till i take that class I guess !
1. When principal (P) is neutral and agent A is risk averse - fixed payment contracts.
and the principal bears all the uncertainty
2. When the P is risk averse and the agent A is risk neutral then : "Sell the project to the agent". i.e. franchise contarcts.
e.g. the agent pays McDonalds to own the franchise - McD gets a fixed payment and the franchise takes all the risk.
However in real life both principal and agent will have some kind of risk aversion and so in real life the contract will be something between the two.
The contracting timeline:
1. principal designs contract
2. agent accepts of rejects contract (IR contract - forms one variable in the lagrangian later)
3. how hard is agent going to work (Incentive Compatibility Constraint) (forms another variable for the lagrangian later)
4. Uncertainty is resolved, i.e. we have some outcome; important point to consider
how do payoffs depend on the outcome of the project - design decision.
5. contract payoffs are made - and happily ever after :)
The timeline allows us to use backward recursion to design optimal contracts.
Lots of weird, torturous equations later :)
IR > 0 => how costly is it to get the agent to just sign ujp
IC > 0 => how costly is the lack of monitoring of information for the principal
and most importantly
t(x)= optimal payment to agent
t(x) depends directly on the likelihood ratio : i.e. the ratio p(i,L)/p(i,h)
{ note here:
derivative of utility of agent t(u(x)) is decreasing
=> as t(x) increases => du decreases => 1/du increases
}
where p(i,h) represents how likely it is to observe result i (state i of the world) if agent worked hard (h)
so if p(i,h)/p(i,l) is small we are much more likely to see results x(i) if effort to high i.e. e=e(h)
the takeaway is since we can't measure effort we measure some quantity that encapsulates effort. here x. so effective you are being paid not for e but for the information encapsulated and communicated by x.
Thus we want to motivate profitable behavior for principal my making working efficiently attractive for agent well - and be able to monitor that in some observable outcome.
For example Varszegi paid his employees at fotex four times the market price in the new hungarian regime - their incentive ? - to keep their jobs by showing Varszegi that they were valuable to his business by working hard. So the profitability of the business encapsulated some information about the work of the employees and Varszegi made a killing. Stock options in companies have the same purpose.
The design question for me is: how many ways can we leverage this in interaction design. How can the interface encapsulate knowledge about user actions and then reflect that to either motivate others or motivate the same user; Are recommender systems at least in part an instantiation of this - or not ? i'll have to wait till i take that class I guess !
Comments