A match brings together two programs written in a custom dialect of Scheme. Each is called with two inputs: the source code of itself and the other. (Giving it its own source code is only a convenience, of course.) It's allowed to run for some standard number of steps ('fuel'). The output should be 'C or 'D within the allotted time. If not, both programs earn 0. Otherwise, standard prisoner's-dilemma payoffs.
In a round-robin tournament, agents get ranked by total score over all matches.
Implementation in Gambit Scheme using its WITH-EXCEPTION-CATCHER primitive; I believe it's portable Scheme otherwise.
(RUN fuel expression environment) is like the standard EVAL but with limited fuel, and returning a list like (unconsumed-fuel result-value). In the scope of this run, (GLOBAL-ENVIRONMENT) returns the environment passed in. The result in the case of an error or fuel exhaustion is defined in terp.scm -- I may want to change all these particulars.
To implement Eliezer's
This enables you to say, "Simulate my opponent, and if it tries to simulate me, see what it will do if it simulates me outputting Cooperate."
rebind RUN to your preferred function in the environment passed to RUN. (For convenience a REBIND function is supplied.)
;; Example agents ;; Trivial cooperator and defector (lambda (me them) 'C) (lambda (me them) 'D) ;; Cooperate with shallow, cooperative agents (first cut) (lambda (me them) (let ((result (run 1000 (list them them me) (rebind 'run run ;TODO: interpose a new RUN (global-environment))))) (let ((remaining (car result)) (value (cadr result))) (if (< 500 remaining) (if (equal? value 'C) 'C 'D) 'D))))
run it's a subset of standard Scheme: (TODO add explanations)
Constant (TODO always expand these out to quotes)
(if exp exp exp)
(lambda (variable ...) exp)
(exp exp ...)
TODO probably nicest overall to include
- IEEE double
- empty list
expand: expand out derived forms, reducing to core forms
(let ((variable exp) ...) exp)
TODO add the usual macros