U-Mart Project＞The Process of Open Experiments

index＞The Process of Open Experiments

The Process of Open Experiments

[PreU-Mart2000]
[Agent type]：Machine
[Conditions] : 15(Session interval: seconds) * 4 (# Sessions) * 60 (# of days)
[Description] :Held to see if the whole U-Mart System would work correctly.

[U-Mart2001]
[Agent types] : Machine and Human
[Conditions] : 15(Session interval: seconds) * 4 (# Sessions) * 60
[Description] : First full-scale open experiment. Participating machine agents gathered and conducted several experiments prior to U-Mart 2001 to collect data for the final evaluation.

Domestic open experiment
Only human agents participate. (Each participant conducts the experiment only one time during the session.)

International open experiment
Only machine agents participate. (Comprehensive evaluation method based on Pareto-ranking concept is used.)

[U-Mart2002]
[Agent types] : Machine and Human
[Conditions] : 10 (Session interval: seconds) * 8 (# Sessions) * 24 (# of days)
[Description] : Many students from Osaka Sangyo University and Chuo University participated and a large-scale human agents’ experiment was held.

[U-Mart2003]
[Agent types] : Machine and Human
[Conditions] : 10 (Session interval: seconds) * 8 (# Sessions) * 30 (# of days)
[Description] : It was held as a demo session of International ISAGA Conference,there were many all comers. New system, U-Mart Ver. 20 was used for the first time.

[U-Mart2004]
[参加形態]：マシン･エージェント＆ヒューマン･エージェント
[設定]：板寄間隔30秒×板寄回数2回×20日
　　　　板寄間隔1秒×板寄回数90回×20日
[特色]：ルールを変更しセッション間の時間間隔や、一日のセッション回数を2種類用意した。インターネットを利用して初めての2地点実施をした。

[UMIE2002]
[Agent type] : Machine
[Conditions] : 4 (# Sessions) per day* 60 (# of days) ... Acceleration experiment
[Description] : Acceleration experiment had been premised and machine agents suitable for such experiments were invited. Comprehensive evaluation method based on Pareto-ranking concept was used.

[UMIE2003]
[Agent type] : Machine
[Conditions] : 4 (# Session) per day * 60(# of days) ... Acceleration experiment
[Description] : Fewer, but "strong" agents participated, reflecting the results of UMIE 2002.

[UMIE2004]
[参加形態]：マシン･エージェント（加速実験）
[設定]：板寄間隔0秒×板寄回数4回×60日
[特色]：各大学から新しいチームが参加した。Classifierエージェントに加えて人工知能技術を利用したオンライン学習型のものが優秀な成績を収めた。

[ Basic rules of competitions ]

Basic rules (Trading rules of U-Mart futures market)

Initial Amount of Cash : 1,000,000,000 yen
Trading unit : Multiplying contract index by 1000 = contract amount
Margin rate : 300,000 yen per trade
Fee: None (0 yen)
Lone limit : 30,000,000 yen
Interest rate of loan : 10%

Rules of real-time experiment (with human and machine agents)

U-Mart2002・U-Mart2003

Trading period: 30 or 24 (virtual)
# Sessions per day: 8・times
Session interval: 10 seconds
Time series data: Distribution kit J30 (experimental data is not released until the experiment starts)
－Human agents and machine agents trade in the same market.
Use of graphical user interface (GUI)
Strategic machine agents participate. Participants bring their own PCs.

【Evaluation method】

Coalition and discussion with team members are allowed. Both personal and team performances are evaluated.

Rules of acceleration experiment (with only machine agents)

U-Mart2002・U-Mart2003

【Market rules】

Trading period: 60
# Sessions per day: 4・times
Session interval: 0

【Summary】
Each agent is scored and ranked total 125 times, with 5 types of experiments (Ex1, Ex2-1, Ex2-2, Ex2-3, Ex3).
－Experiments with different agents compositions.
－4 types of time series for spot prices(up, down, reverse, oscillate) are used.
－Conduct experiment under various market conditions.

【時系列】
上昇･下降･反転･振動の４系列。下の図は実際に使われた時系列である。

	上昇系列	下降系列	反転系列	振動系列

【実験】
（エージェントの組み合わせによって以下の３種類の実験を行う）

EX1）各応募エージェント１匹＋標準エージェントセット20個体
EX2）全応募エージェント＋標準エージェントセット20個体
EX3）予選通過エージェント＋標準エージェントセットからランダムに半分選択

【Evaluation method】

[Comprehensive ranking] : Rank based on the following four criteria and using Pareto-ranking concept.
[Maximum gain] : Evaluate based on the maximum gain amount throughout the session.
[Average gain] : Evaluate based on the average gain amount throughout the session.
[Profit count] : Number of experiments finished in the black throughout the session.
[Bankruptcy count] : Number of experiments terminated by bankruptcy throughout the session.

【Experiment procedure】
Pattern 1 (for EX1, EX2, EX2-2, EX2-3)

Initialization --> Repeat same sequence 50 times
Repeat 4 patterns (up, down, reverse, oscillate) total 200 times
Evaluate each sequence and total of all the sequences (for 5 types)
Rules of real-time experiment (when human and machine agents coexist in the market)
Rules of acceleration experiment (when only machine agents exist)

・パターン２　（Ex3の場合）

a) 全エージェント（参加エージェント＋標準エージェントセット）からランダムに半分のエージェントを取り出す
b) ４種類の時系列からランダムに一つ選び出し実験する。
上記　a), b) を1000回繰り返す。

【ジェントルマン・ルール】
エージェントは、積極的に取引に参加する事