2.7 Experimentation
3.1.3 Decision making algorithms of software agents
3.1.3.6 Other strategies
Offer generation based on other than the above mentioned approaches are often very simple ’dummy’ strategies employed to verify the performance of strategies based on evolutionary com- puting, imitation, learning, or time. Such simple approaches continuously make concession steps of a predetermined magnitude (Park and Yang, 2004; Zeng and Sycara, 1998), by reducing the gap in the demands by some fixed proportion (Somefun et al., 2004, 2006), or simply take ran- dom steps (Winoto et al., 2005; Henderson et al., 2003). In some simulation studies, however, such ’other’ strategies are also quite sophisticated ones but build on approaches different from those mentioned above. Such alternative sophisticated strategies are not so frequently used and we therefore only briefly enumerate them: hill-climbing and annealing (Klein et al., 2003), least- cost-issue concession (Wasfy and Honsi, 1998), Zeuthen-risk strategy (Wollkind et al., 2004), case experience-based strategies (Paurobally et al., 2003), or competition-based strategies (An et al., 2006).
In eight studies evolutionary computing are used to determine the offer generation of the soft- ware agents. While most of them use simple binary chromosomes with offers and thresholds (sequential threshold rules), Tu et al. (2000) and van Bragt and La Poutre (2003) encode finite state machines in form of chromosomes for evolutionary computing. 14 studies employ imitating strategies, which in most cases acted purely reciprocating. However, some variations of imitating strategies also employed include those proposed by Krovi et al. (1999), which are: (i) recipro- cating strategies that fully reciprocate the behavior of the opponent, (ii) cooperative strategies that more than match the opponents concessions, and finally (iii) exploitative strategies that
3.1. Current achievements in simulation of automated negotiation 63
concede less than the opponent. Trade-off mechanisms employing fuzzy-similarity measures, Pareto-search, or other similarity measures to make offers close to those of the opponent were used in ten studies – such trade-off mechanisms for multi-issue negotiation problems find in- creasing attention in recent years. Trade-off mechanisms typically are combined with some rules about how to determine changes in the demanded utility level, if no trade-off offers (with same utility) are left to propose, to form the final strategy of the agent. In nine studies learning agents with a variety of learning techniques were investigated. The employed learning techniques range from Bayesian network updating (Zeng and Sycara, 1998) over neural networks (Rau et al., 2006; Park and Yang, 2004) and pattern recognition (Lin and Chang, 2001) to Q-learning (Cardoso and Oliveira, 2000). Another 14 studies rely on time-based concession functions to determine offers.
As many studies derive strategies by a combination of approaches or compare different types of strategies in their simulation studies, the overall number of types of strategies used for offer generation in the studies not matches the number of studies reviewed. These comparisons lead to interesting results, which are informative for the design of decision making algorithms for software agents. Zeng and Sycara (1998) show that learning agents can exploit and outperform simple agents with continuous concession strategies. Deveaux et al. (2001) show that agents that hold a model and an expectation about the opponent’s strategy in mind and adapt their strategy according to the information they gathered during the negotiation outperform time- based strategies. van Bragt and La Poutre (2003) find that strategies embodying a finite state machine improved by means of evolutionary computing can exploit time-based and imitating strategies, and Tu et al. (2000) state that strategies based on evolutionary computing with chromosomes encoding finite state machines do not perform significantly better than those with chromosomes encoding sequential threshold rules.
In addition to comparisons of different software agent strategies in terms of their performance, few studies also compare the results of software agents in automated negotiation (i.e. simula- tion results) to the results predicted by game theory (i.e. analytical solutions), to unsupported human negotiation experiments or experiments where humans are supported by negotiation sup- port systems, or even let humans negotiate against software agents in negotiation experiments. Gerding et al. (2003) show that agents based on evolutionary computing reach the results pro- posed by game theory for rational agents in bargaining games. They propose and accept extreme take-it-or-leave-it offers at the final round of games with a finite number of rounds and reach results near the analytically derived subgame-perfect equilibrium (Rubinstein, 1982) in the first round when the game is characterized by a small probability of negotiation break off after each round. The results for human agent versus software agent comparisons, however, are somewhat disappointing for the field of automated negotiations. Oliver (1996) compares the performance of his software agents ,based on evolutionary computing, to the performance of humans in exper- iments for negotiations reported by Raiffa (1982) and Rangaswamy and Shell (1997). He finds that his software agents matched the performance of unassisted negotiators but are outperformed by humans that use negotiation support systems in integrative negotiation problems. Similarly Goh et al. (2000) find that in distributive negotiation problems there is no significant difference in performance for humans negotiating via a messaging system, via a negotiation support sys- tem that provides analytical support, or software agents that use simple continuous concession functions. In integrative negotiation problems ,however, negotiation support system users out-
perform messaging system users, which in turn outperform software agents. Bosse and Jonker (2005) use preferences elicited from human users to simulate automated negotiations and also find that the performance of software agents based on simple concession functions matches the performance of human negotiators if the agents face opponents of the same type – i.e. computer versus computer and human versus human. However, in an experiment where humans negotiated with software agents, that use the preferences elicited from humans as input, the human subjects reached significantly better outcomes than their opponent software agents.