THE MOVE-DECISION STRATEGY OF Indigo

Abstract

This paper describes the move decision strategy of Indigo. By using the example of Indigo, the paper shows that the move decision process of a Go program can be very different from the processes used in other games with lower complexity than the complexity of Go, even if the basic modules are conventional (move generator, evaluation function and tree search). Indigo uses them in a specific way, adapted to computer Go, which may be of interest for researchers on other mind games as complex as Go. The evaluation function can be "quick", "slow" or "strategic". It may or may not include local tree search. The move generation brings about different kinds of moves : "urgent" moves, "life and death" moves and "calm" moves. Urgent moves are statically qualified with a global urgency. A two-player quiescence search verifies that the urgent move does not decrease the position evaluation. Calm moves are used within two-player selective global search at a very low depth. Besides, Indigo also uses single-agent search to refine the strategic importance of goals. Lastly, Indigo chooses either the calm move, the life and death move or the urgent move to be the global move.

Monte-Carlo Go Reinforcement Learning Experiments

“…Moreover, the bad moves are never definitely eliminated from the process. In 2002, our experiments carried out with Bernard Helmstetter, a doctoral student under Tristan Cazenave's supervision at Paris 8 University, showed that, on 9x9 boards, pure MC programs ranked on a par with heavily knowledge based programs such as Indigo2002 [35]. Given the architectural difference between these programs, that result was amazing.…”

Associating domain-dependent knowledge and Monte Carlo approaches within a Go program

“…Indigo [6,5] is a classical go program based on tree search [8] and on extensive knowledge [9]. For instance, territories and influence are modelled by means of the mathematical morphology [7].…”

“…On 13x13 boards, Olga obtains her best winning percentage. However, Olga(pseudo = true, N s = 1) corresponds to the urgent method [8] of Indigo2002 selecting one move without verification. Its level is necessarily inferior to the one of Indigo2002 that uses a calm method in addition to the urgent method with verification [8].…”

Monte-Carlo Go Developments

“…OLGA and OLEG are far-fetched French acronyms for "ALeatoire GO" or "aLEatoire GO" that mean random Go. OLGA was developed by Bouzy (2002) as a continuation of the INDIGO development. The main idea was to use an approach with very little domain-dependent knowledge.…”

“…We used the 3.2 version released in April 2002. INDIG02002 (Bouzy, 2002) is another knowledge-based program whose move decision process is described in Bouzy (2003). OLGA means 0LGA(Depth=1, rd=1, ae=0.2) using PP and not the all-moves-as-first heuristic.…”