modelling self-organisation of oligopolistic markets using genetic programming
TRANSCRIPT
Modelling Self-Organisation of Oligopolistic Markets Using
Genetic ProgrammingEdmund Chattoe
Department of SociologyUniversity of Oxford
[email protected]://www.sociology.ox.ac.uk/people/chattoe.html
Plan of the Talk
• The research question• Genetic programming• Results of the simulation• Qualifications and conclusions
Oligopoly Pricing• How do firms set prices in a complex
environment?• Simplify by making it a game or assuming
(“gifting”) lots of common knowledge• Third approach is adaptive but this is “too difficult” for simple adaptation
• A possible solution is evolutionary learning: firms adapt by an ex ante undirected mechanism and are selected according to ex post success
• Descriptive and instrumental applications
The Intellectual Appeal of Evolution• Driven by heterogeneity• Open ended: actors don’t need to know
objective function (if there is one)• Works on minimally effective strategies
using relative success• Analogous to situation of firms?• Already observed to produce stable self-
organised heterogeneity in ecosystems
A Brief History• Marshall and the representative firm• Alchian
– Outcomes not intentions– Genotype is firm practices– Phenotype is firm behaviour (and structure)
• Nelson and Winter– Fixed decision rules (link to last week)
• Dosi et al. (1999)
Evolutionary Algorithms• Genetic Algorithms and Genetic Programming• Population of solutions represented as data
structures (lists and trees): Travelling Salesman• Population enrichment by selection (GENITOR
versus HOLLAND)• Genetic operators: Crossover, inversion and
mutation• Shared representation/interpretation of imitation• Justification for using “hill climbing” operators• Exploring full space of grammatical strategies
A Typical GP Price String+
3/
OP1
+ 2
OP2
The Dosi et al. Model• Candidate prices are small set of GP strings• Firms set price probabilistically based on
accumulated (but bounded) profits of candidates• Demand determined by “market” price and
allocated by current market share• Market share updated via set and market prices• Profits are accrued to firm (and strategies)• Firms with losses/minimal market share replaced• New candidates may be generated
Modifications to Dosi et al.• Note: No ability to predict effects of pricing• No exogenous “restarts”• Behavioural interpretation of small set of competing
price strategies• Fair trial assumption: priority to untried strategies
and no profit “inheritance”• Genitor rather than Holland architecture: much
smaller number of candidates• Consistent treatment of imitation in new and
surviving firms
The Operators• Crossover: Take two trees, identify “legal” cut
points and swop “tails”• Mutation: Take one tree and identify “legal” cut
point for new randomly generated tree.• Some completely new trees• Other possibilities?• IF NOT AND OR > < = + - % *• OMP, OMD, OP x OUC, CUC, OS, integer• http://users.ox.ac.uk/~econec/thesis.html
Results: Checking Replication• Replicating the Dosi et al. simulation from
the published description: not an easy task• Check the simulation by learning the
monopoly price• Noise reflects experimentation: wasteful but
only if you know the underlying properties of the search space
• Three replication runs per result: not reproduced here
Monopoly Learning
Results: Dosi et al.
• Main Dosi et al. result is evolution of price following and “cost plus” pricing
• This appears to be sensitive to assumptions made about relatively large and variable unit costs
• Use of “elasticity measures” to confirm this result in a more quantitative way
Dosi et al. Replication
Cost Plus Pricing
Price Following
Much Lower Unit Costs
Results: Salience and Co-ordination• With low and fixed unit costs, market is “speculative”: no fundamentals
• Salience is a property that makes some choices stand out for reasons irrelevant to their payoffs
• 1 2 3 4 5 6 7 8 9 10• Beliefs can make choosing salient options rational• Offer firms “common” (but non optimal) terminal
without multiples• New mode of explanation: neutrality about whether
terminal is used in real strategies or not
Speculative Market: Fixed Unit Cost
Co-ordination Through Salience
The Role of Expectations• Recall the Dosi et al. findings involve “backward
looking” strategies• Feedback to market share produces strong
tendency to monopoly: prices are relatively stable but not co-ordinated and thus not sustainable in the long run
• Firms assess strategies on profit only• Adding a simple linear expectations operator
ensures both stability and co-ordination
Stable But Uncoordinated Market
Market Sustainability
Expectation Formation Terminals
A Mechanism of Tacit Collusion• In a duopoly both firms gain from co-ordinating on
the same price: splitting the market• Both leader and follower might lose profit (via
market share) if one “defected” from this• Expectation terminals give a mechanism for tacit
collusion• The leader “trails their coat”• The follower can use expectations on leader prices
that do not change too sharply• An emergent (and sustainable) property?
Tacit Collusion?
Sustainable Market Shares
Three Firms with Expectations
Changing Firm Goals• Introduce possibility of market share goal• Populations with mixed goals• Profit maximisation doesn’t drive out other
goals (contra Friedman)• “Pure” market share maximisation leads
invariably to monopoly in mixed goal markets
What Next?• Better data collection: ethnographic,
experimental, participatory• More effective sensitivity analysis:
particularly challenging Dosi et al. results• Problem of behavioural interpretation• Experimental programme: Diverse strategy
sets, multiple control variables, different operators and terminals
• Better representation?
Conclusions• GP is a powerful adaptation mechanism• Evolutionary algorithms may also be used
as descriptively plausible models for adaptation in economic systems
• Augmenting the grammar of a GP offers a less “judgmental” experimental approach to firm capabilities
• GP models might contribute to “standard” theory of the firm