Overview
This is a semester project for Master student aimed at investigating collusive
outcomes when autonomous agents learn to bid in electricity markets. The
project will primarily involve numerical simulations, but requires the student to
have a good grasp of theory as well.
Outline
Electricity markets are organized as auctions where producers place bids for
supplying electricity. Based on these bids, the market operator selects a price and
schedules the quantity of the electricity to be sold. The electricity market is
cleared repeatedly for each interval of the day, and market operators publish
market clearing data for each clearing interval.
If electricity markets have a small number of producers, then these producers are
said to have market power. Producers with market power can inflate their
electricity costs to get higher prices and profits in electricity markets, at the
expense of consumers. Thus, it can be in the interest of individual producers to
coordinate production as a single entity. To mitigate this, regulators break up
large power companies (divestiture) to increase competition and increase social
welfare. Regulations also forbid explicit collusions, where companies agree to
coordinate their bids to increase prices (cartels).
The recent years have seen an increase in the deployment of automated bidding.
Recent works [1,2] have shown that even simple reinforcement learning
algorithms used by bidders can result in implicit collusion. This means that
bidders need not explicitly communicate to coordinate their bids, they can
implicitly infer others’ behavior from the market clearing data published by the
operator, and then learn to collude with each other. This is known as algorithmic
collusion.
There have been recent works on algorithmic collusion in different domains such
as online marketplaces [3], rental platforms [4]. However, research related to
electricity markets [2] is limited and ignores realistic aspects of the grid, such as
stochastic load and generation, network constraints, and risk-aversion of
producers.
The project will consist of two parts. The first part will involve creating a market
simulator where market outcomes are compared for different learning
algorithms: tabular reinforcement learning [1,2], bandit learning and best-
response dynamics [5,6]. The second part will then involve a numerical study on
how modifying information disclosure (market clearing outcome data) affect
collusion outcomes.
[1] Emilio Calvano et al. “Artificial Intelligence, Algorithmic Pricing, and
Collusion”. In: American Economic Review 110.10 (Oct. 2020), pp. 3267–3297. issn:
0002-8282.
[2] Ibrahim Abada and Xavier Lambin. “Artificial Intelligence: Can Seemingly
Collusive Outcomes Be Avoided?” In: Management Science 69.9 (Sept. 2023), pp.
5042–5065. issn: 0025-1909, 1526-5501.
[3] Francesco Decarolis, Maris Goldmanis, and Antonio Penta. “Marketing
Agencies and Collusive Bidding in Online Ad Auctions”. In: Management Science
66.10 (Oct. 2020), pp. 4433–4454. issn: 0025-1909, 1526-5501.
[4] Sophie Calder-Wang and Gi Heung Kim. Algorithmic Pricing in Multifamily
Rentals: Efficiency Gains or Price Coordination? SSRN Working Paper. Available at
SSRN: https://ssrn.com/abstract=4403058. Aug. 2024.
[5] Arega Getaneh Abate et al. “Learning to bid in forward electricity markets
using a no-regret algorithm”. In: Electric Power Systems Research 234 (Sept. 1,
2024), p. 110693. issn: 0378-7796.
[6] Orcun Karaca et al. “No-Regret Learning from Partially Observed Data in
Repeated Auctions”. In: IFAC-PapersOnLine 53.2 (2020), pp. 14–19. issn: 24058963