Camerer and Ho - EWA Flashcards
(9 cards)
What is EWA learning?
A model developed by Camerer and Ho that hybridises and generalises reinforcement learning and belief based learning approaches to describes how people learn in strategic situations through experience.
What two learning models does EWA unify?
Reinforcement learning (where strategies that worked well in the past are more likely to be chosen again) and belief based learning (where players form beliefs about others’ strategies and best respond)
What are the key variables tracked in the EWA model?
1) Attractions for each strategy (A), which determine choice probabilities
2) Experience Weight (N), which measures the strength or reliability of attractions; and
3) Choice probabilities based on attractions
How is the experience weight updated in the EWA model?
N(t) = p.N(t-1) + 1
Where p determines the rate at which previous experience is discounted.
What are the key parameters in the EWA model and what do they represent?
1) delta - imagination parameter. The weight placed on forgone payoffs.
2) Phi - the decay parameter. The discount factor for previous attractions.
3) row - the experience discount. Discount factor for past experience.
4) lamda - response sensitivity. How strongly attractions map to probabilities.
How does EWA reduce to simple reinforcement learning?
When delta is equal to 0 (only realised payoffs matter), row is equal to 0 (so no expereince accumulation) and phi is equal to 1 (no decay of previous attractions)
How does EWA reduce to weighted fictitious play (a belief based model)?
When delta = 1 (realised and forgone payoffs weighted equally) and row = phi (experience and attractions decay at the same rate)
What does a high alpha parameter indicate about a player’s learning process?
Means attractions decay slowly, indicating a player with good memory. who gives substantial weight to past experiences when making current decisions.
What does a low lamda parameter indicate about a player’s choice behaviour?
A low lamda indicates more random choice behaviour when differences in attractions have less influence on choice probabilities (more exploration or error)
