Evaluation function

Criteria

Efficient computationally
Accurate representation of utility
- $Utility(loss,player) \le Eval(state,player) \le Utility(win,player)$
Agrees with terminal states
- $Eval(state,player) = Utility(state, player)$

Single value return that estimates the proportion of state with each outcome
- Calculate expected value of state with sum of utility and corresponding reward
Weighted linear function
- Weights of each feature summed which forms the estimation of a state (Linear combination calculation)
- $Eval(state)=w_1f_1(state)+w_2f_2(state)+...=\displaystyle\sum_{i=1} ^{n}{w_if_i(state)}$
- Strong assumption made that each feature is independent of values of other features
  - A pair of bishop might be worth more than twice the value of a single bishop in chess
- Non-linear combinations can be made to resolve this issue

Links: