Title: On Learning in Policy Space by Oligopolists in Electricity Markets
1On Learning in Policy Space by Oligopolists in
Electricity Markets
- by
- Steve Kimbrough
- Fred Murphy
2Organization of Presentation
- Bidding process in day-ahead market
- Supply function equilibria
- Agent-based modeling
- Bid representation
- Results
- Conclusions
3Day-ahead Market
- Each firm bids a step-function supply curve
- LSEs offer demand curves
- The supply and demand curves are put into an LP
with transmission capacity constraints and the LP
solved for the equilibrium
4Our Day-ahead Market
- Demand curves are continuous
- No transmission constraints
- Firms bid step-function supply curves
- Combine individual supply curves into single
supply curve - Find market equilibrium, where curves cross
5Alternative models of Equilibria
- Cournot - bid quantities
- Problem utilities bid prices as well
- Bertrand - bid prices
- Problem oddly discontinuous behavior
- Add capacity game and same as Cournot
- Supply function bid a supply curve
6Supply Function Equilibria
- Klemperer and Meyer (1989), Green and Newbery
(1993) - Literature uses continuous curves to derive
results - Models have multiple equilibria unless severely
restricted in functional form and/or domain
7A Problem with all of the Standard Models
- Firms are part of a community
- The models assume myopic optimization and pure
non-cooperative behavior - There are institutions where firms can talk
without breaking the law - Leading companies look to the interests of the
industry as well as the firm
8Agent-based Modeling
- Two approaches
- Agents are decisions/policies and fittest survive
- Agents try different decisions and mostly go with
the decisions that have better outcomes - Our approach is to
- Give agents alternative objective functions
- Let agents make decisions based on the values in
the assigned objective function - Evaluate outcomes based on firm profitability
9The Model of the Market
- Player objective functions
- Firm profitability
- Industry profitability
- Industry profitability subject to a fair share
constraint - Supply functions
- n plants, n steps
- n plants, n1 steps
10The Model of the Market, cont.
- Episode - one round of play
- Players randomly adjust prices and quantities
around trial values on steps - Epoch - a collection of episodes
- Players evaluate the outcomes from the random
trials using the assigned objective function - They adjust the trial values in the direction of
the increased objective function
11Monopolist, N
File elec-Own-N.png
12Monopolist, N1
File elec-Own-N1.png
13Duopoly, Own-Own,N-N
Players find Cournot outcome. High-bidding
player is exploited.
File elec-Own-Own-N-N.png
14Duopoly, Own-Own, N1-N1
Players find Cournot-plus outcome. Neither
player is exploited.
File elec-Own-Own-N1-N1.png
15Duopoly, Ind-Ind, N1-N1
Players individually maximize industry profits
and jointly find monopoly outcome.
File elec-Industry-Industry-N1-N1.png
16Duopoly, Ind-Own, N1-N1
Players find near-monopoly outcome. Cooperative,
Industry Returns player, is exploited.
File elec-Industry-Own-N1-N1.png
17Duopoly, IndOwn-Own, N1-N1
Players find Cournot-plus outcome. Cooperative,
Industry Returns, s.t. Own Returns player, is
not exploited.
File elec-IndustryOwn-Own-N1-N1.png
18Duopoly, IndOwn-IndOwn, N1-N1
Players find monopoly outcome. Neither player is
exploited. (But the customers are.)
File elec-IndustryOwn-IndustryOwn-N1-N1.png
19A Pattern
- Classical results hold for learning agents who
are very myopic. - Agents learn to collude tacitly by being less
myopic and trading off exploration and
exploitation, tilting more towards exploration.
(Exploring rationality) - Simple tilts towards exploration are subject to
exploitation. - There exist simple constraints on exploration
that greatly reduce exposure to exploitation. In
this mode agents may mutually achieve tacit
collusion safely.
20Conclusions
- In repeated play the structure of the supply
curves can be exploited to find the Pareto
optimum - At the same time it is possible for a firm to
protect itself from self-serving players - To get its fair share, each player has to make
the other see the consequences of its actions on
aggregate demand in its choices of prices and
quantities