next up previous
Next: SSA-enabling Instructions Up: Semantics of Instruction Heads Previous: ``Introspective'' Instructions


Bet Instructions

Bet!( $x_1,x_2,y_1,y_2,c,d$): Select a value $c' \in \{ 0,
\ldots, n-1 \}$ with probability $ \frac{{\sc Left}_{IP+5,c'}}
{\sum_k {\sc Left}_{IP+5,k}}.
$ If $c' > n/2$ then set variable $c \leftarrow 1$, else $c
\leftarrow -1$. Select a value $d' \in \{ 0, \ldots, n-1 \}$ with probability $ \frac{{\sc Right}_{IP+5,d'}}
{\sum_k {\sc Right}_{IP+5,k}}.
$ If $d' > n/2$ then set variable $d \leftarrow 1$, else $d
\leftarrow -1$. If $c = d$ then exit. If $\cal S$$_y = \cal
S$$_x $ then give reward $c$ to LEFT and reward $-c$ to RIGHT. Otherwise give reward $-c$ to LEFT and reward $c$ to RIGHT. Set $\cal S$ $_{7} \leftarrow c$ (surprise rewards become visible to the system in the form of inputs).

Comment: arguments $x,y$ are indexed by $IP+1, IP+2,
IP+3,IP+4$, while arguments $c,d$ are chosen according to module columns indexed by $IP+5$. See Section 2 for Bet!'s interpretation.



Juergen Schmidhuber 2003-03-10


Back to Active Learning - Exploration - Curiosity page