Experience-Weighted Attraction (EWA) learning method
Learning is a process of acquiring new, or updating existing, knowledge, behavior, skills, values or preferences, and learning is not only from one’s own experiences but also from the experiences of others (Crawford, 1995). Economists have developed three prominent models to describe the learning processes: Reinforcement Learning (RL), Belief Learning (BL) and EWA Learning (Thomas, 2006). In the RL model, behavior that has led to success in the past and present will have a higher frequency of occurrence in the future, while actions that have caused failure will be less likely to be chosen in the future (Jasmina and Alexander, 2010). The BL model addresses how individuals formulatebeliefs or take actions based on a weighted average outcome of strategies that other people have adopted in the past (Camerer et al., 2004). However, the BL model ignores the role of individuals’ past experience in formulating new strategies, while the RL model does not account for comprehensive information about strategies that are not selected (Jasmina and Alexander, 2010). Camerer and Ho (1999) built a EWA learning model, which in essence integrated both RL and BL models. The EWA model is capable of capturing individuals’ own experiences and allowing people to adjust their strategies based on changing contextual factors (Camerer et al., 2004). This model not only considers previous land-use strategies that farmers used before the completion of the GFG program in 2006, it also estimates potential change of their land-use strategies after the completion of the GFG program.
This study uses the EWA model to simulate possible change in land-use strategies among different farmer groups. The EWA model used in the study is expressed in Equations (1), (2) and (3). The two primary outcome variables included in the model are the following: N(t), referring to the weight of past land-use strategies (i.e., experience weight); and Aji(t), standing for the attraction probability of land-use strategy j to farmer groupi at the end of the first stage of the GFG program in 2006.
t≥1 (1)
(2)
Whereρ is the depreciation rate of the previous land-use strategy,N (t-1), that might be utilized in the subsequent period. The parameter δ determines the extent to which hypothetical evaluations will be used to compute the attraction probability. If δ = 0, strategies that were not selected in the past will not be input to the model; in this case the model evolves into the RL model. If δ = 1, strategies which were not selected in the past will be fully integrated into the hypothetical evaluations of the model. In this study δ = 1. The parameter Φ is another depreciating rate of the attraction probability. Si(t) refers to actual land-use strategies of farmer group i at t. S-i(t) is a vector matrix of actual land-use strategies of other farmer groups intime t.
The probability of land-use strategy j that might be selected by farmer groups after the completion of the first stage of the GFG program at a later time t+1 (i.e., post 2007) is expressed in Equation (3):
(3)
The initial values of N(0) and A(0) can be assumed on the basis of learning from past similar land-use strategies. This study employs the initial values from the modelling results of experimental studies of Camerer et al (2001), in which ρ=0.935, φ=0.986, and N(0)=1/(1-ρ)=15.385. We assume that Aji(0) is the income weight of a land-use strategy selected by farmer group i at the initial phase. The attraction probability of land-use strategy Aji(1) refers to the likelihood of a land-use strategy collectively selected by all farmer groups during the 1999-2006 period, while the attraction probability of land-use strategy Pji(2) refers to the likelihood of a land-use strategy to be selected by all farmers following completion of the first stage of the GFG program in 2006.
Weutilise the experimental values from the study by Camerer et al. (2001) as the initial model values in this study primarily for two reasons. One is that the parameters defined by Camerer et al. (2001) capture changes in attraction probability of possible responses to policy change. The other reason is that there have been no experimental or initial values to date tailored to an analysis of farmers’ land-use strategies, despite a growing number of studies in the field of economics using the EWA model.