Generalizing from Examples

• previous algorithms make no attempt to estimate the Q value for unseen state-action pairs, unrealistic in large or infinite spaces or when the cost of executing actions is high
• substituted ANN for the table lookup and use each update as a training example
• A more successful alternative is to train a separate ANN for each action using state as input and as output
• Another common alternative is to train one network with state as input and with one output for each action
• the convergence theorems no longer hold!!