[Game Theory] The Price of Participation: How Rewards Drive Crowdsourcing Engagement

TL;DR

In the high-stakes world of software crowdsourcing, how much should an outsourcer pay to ensure enough developers join the fray? This paper applies Nash Equilibrium and a novel Sequential Pairwise Challenge model to predict the "Expected Number of Participants" based on reward size. It reveals that the presence of a "superstar" developer fundamentally changes the ROI of rewards, requiring nearly 4x the prize money to achieve full participation compared to a field of average peers.

The "Winner-Take-All" Dilemma

Software development is increasingly shifting toward crowdsourced platforms like TopCoder and AppStore. Most of these contests operate on a winner-take-all basis. For a participant, the decision to join is a gamble:

• The Cost: Time and effort ($c_i$).
• The Benefit: The reward ($R$) multiplied by the probability of winning ($W_i$).

If the expected payoff is negative, the developer quits. The outsourcer’s challenge is to set $R$ high enough to attract talent without overpaying. Prior work treated participants as homogeneous units, but this paper acknowledges that skill asymmetry (how likely A is to beat B) is the real driver of participation.

Methodology: Modeling the "Battle Royale"

The authors model the contest as a non-cooperative game with $n$ candidates. Each candidate has two strategies: $\{Join,Quit\}$.

1. The Payoff Function

The utility of a participant $C_i$ is defined as: $p_i(s)=R\cdot W_i(s)-c_i$ Where $W_i(s)$ is the probability of winning. But how do you calculate $W_i$ when multiple people join?

2. Sequential Pairwise Challenges

The authors propose that the final winner is determined through a series of matches. By calculating the expected winning probability across all possible permutations of these matches, they derive a robust $W_i$.

Figure 1: The algorithm used to recursively compute the winning probabilities for any subset of participants.

Experimental Insights: Three Types of Crowds

The researchers tested their model using the Gambit software to find Nash equilibria across three scenarios:

Case A: The Level Playing Field (Average Capabilities)

When candidates have similar skills, participation scales linearly. As the reward $R$ increases, more developers jump in. In their test with 7 players, full participation occurred at $R=180$.

Case B: The Superstar Effect (One Super Strong Candidate)

This is the most critical finding. If one candidate has a 90%+ win rate against others, the participation curve becomes a ladder.

• The Psychological Barrier: Average players realize their chance of winning is slim.
• High Cost of Entry: To get 7 participants, the reward had to be jacked up to $R=670$ (versus 180 in Case A).

Figure 2: The "Ladder" effect in participation when a dominant player is present.

Case C: Normal Distribution (The Real World)

Using a Gaussian distribution for skills, the results sat between the two extremes. Full participation was reached at $R=340$.

Critical Analysis & Conclusion

This paper provides a rigorous mathematical bridge between incentive design and game theory.

Takeaways for Platform Designers:

1. Transparency Matters: Knowing who else is in the contest (the "Participants list") allows candidates to calculate their Nash Equilibrium, but it may discourage them if a superstar is identified.
2. Reward Scaling: If you know your platform has "power users," you must be prepared to offer significantly higher rewards to maintain a healthy ecosystem of secondary participants.

Limitations: The current bottleneck is computational. Finding all Nash equilibria for games with a large number of players ($n>20$) is NP-hard. Furthermore, the model assumes players have complete information about their opponents' winning probabilities—a rarity in the real world.

Future Outlook: The authors suggest moving toward multiple-winner models to see if spreading the reward pool can counteract the "Superstar Effect" and encourage participation at lower total costs.