Trading in OOTB2006

[Bristolduke]

At the risk of generating another "Oh my god it's terrible" thread, I was curious if anyone as "investigated" trading. I have seen, on different posts, how bad the AI is in areas. As an expert in AI development (sorry I'd have to kill you to discuss), I want to point out the enormous challenges in this type software. Independent thought via software, vice a massive rules matrix is difficult, but of course you must first agree on the intelligence you are simulating (no barbs intended here). Thu sI would like to focus this discussion on trading in one league, I believe from there we can discuss the multiple leagues. I also expect this will impact extensions. Knowing how difficult it is to test AI software, if there are bugs, and I'm sure there are, we will need lots of data to help Marcus and others.

I did a simple test using 2001 historical season. Three runs were may, one very low trading frequency, one average, one very high. Average difficulty and neutral preference. On average 61 trades were made, on very low 23 trades were made, and on very high 88 trades were made.

Some early returns were that the Reds always traded, Michael Tucker, Hal Morris and Dmitri Young early. All three are "expendable" in the sense of Casey will play first, Young is listed at 3B where Aaron Boone sits and the Reds have several OF. They also need catching and pitching and all three trades netted these items.

So far so good. In browsing some of the trades I found this one on the average setting.

Traded 27-year old left fielder G. Anderson and $3,000,000 in cash to the Atlanta Braves, getting 36-year old shortstop W. Weiss in return.

Now this looks dumb. However upon further review, Anaheim was in need of a SS and had OFs to deal. They also have on their team setting a very strong (one notch from max) preference for veterans on their roster. Now is the trade still a good one, well the point of view has changed. We may not like it, but the AI "may" be working within parameters. Anaheim made 5 trades in the season 4 invloved younger players for older players (the 5th was for the same age)


Remember Artificial Inteligence is not "perfect intelligence" nor "stadic/predictable intelligence". It is simulating independent thought along a defined set of knowledge based parameters.

I haven't investigated all the parameters yet, but besides the three settings in game setup, both the team and the GM have settings for each team. (It appears that you can't change the individual coaches preference, but you can the team - thus making it different than the GM).

In my tests, Jan, Jun and Jul were popular months, indicating league standing and perceived weaknesses for each team is a factor. I would also believe the financials are factors. An open question is the owner a factor?

As I indicated I would appreciate anyone else's thoughts investigations. I will contiue to work the analysis as time permits.

[Raidergoo]

Could use a thoughtful trade critic around here. Please post what you think on trades.

The AI needs to make this kind of deal

Deals that look right on paper and at the time, but when seen throught the lens of history are utter disasters.

[Bristolduke]

You would pick that one. As a life long Reds fan, I still have nightmares.

[Raidergoo]

The AI must make this deal too.

November 29, 1971

Cincinnati Reds acquire Joe Morgan, Jack Billingham, Denis Menke, Cesar Geronimo and Ed Armbrister from the Houston Astros for Lee May, Tommy Helms, and Jimmy Stewart

[Arlie Latham]

That is a bad trade.


Jimmy Stewart was well beyond his prime. His MVP seasons of It's a Wonderful Life, Mr. Smith Goes to Washington and The Man Who Shot Liberty Valence are far in the rear view mirror by 1971.

[Bristolduke]

A brief update and a couple of thoughts.

Could use a thoughtful trade critic around here. Please post what you think on trades.

The AI needs to make this kind of deal

Deals that look right on paper and at the time, but when seen throught the lens of history are utter disasters.

The AI must make this deal too.

November 29, 1971

Cincinnati Reds acquire Joe Morgan, Jack Billingham, Denis Menke, Cesar Geronimo and Ed Armbrister from the Houston Astros for Lee May, Tommy Helms, and Jimmy Stewart

For these type of trades (or any actually), there are two different aspects. 1) Would the AI accept such a trade and 2) Would the AI initiate/offer such a deal. The first is an evaluation mechanism and able to create. The second is significantly harder since you are asking the AI to be "very creative" but not frequently creative, it may however escalate into the deal through negotiation (a little easier).


And of course what type of AI should we have? I'm sure there will be many opinions which will lead to "customization".

I ran two more tests and realized I need to reset my base. The test were average frequency with one on very easy difficulty and 1 on very hard difficulty. The number of trade were close (orig baseline was 61) as very easy had 57 and very hard had 61.

I was analyzing trades with respect to Raidergoo's question and found this anomaly. In the average test 28 of the 61 trades included cash and there were no multi-player deals (the other two tests were similar 30 of 57 involved cash on the easy difficulty and 35 of 61 involved cash on hard.). This led me to fear that since I have minors turned off those teams looking for prospects can't find them and deal. Thus are they settling for cash, not trading or setting for an equivalent player.

While I wanted the tests to be simple initially, I think minors have to be enabled. It also means Imay need to run the sim longer in evaluating the prospects to see if the potential were there.

Two patterns did exist. The maximum cash was 3 million (it occurred frequently and did not appear to be entirely based on need (e.g Player A and $3M to the Yankees for Player B). Second, was the fact that there were no multi-player deals. I know I was offered some in my other sims, so again it may be the lack of minors is limiting.

In restarting, some tests, I will look at the following configurations.

Start in 1996, that will give a better chance to look at the prospects.
Set all the teams to neutral for all strategies. (Initially this will take that variable out of the equation).
Set Trade preference to Neutral.

Then test Average settings for frequency and vary difficulty. Repeat for very high and very low frequency.

[Raidergoo]

Have you checked out the Anderson Utilities yet?

Anderson Utilities 2006

A new feature he is developing seems perfect for any hindsight review you want to run.

[Bristolduke]

Work will interfere somewhat here, but based on the lack of posts, maybe this isn't the issue I had assumed.

Anyway, I have simmed 1996 three times (twice with average frequency and average difficulty and once with average frequency and very hard difficulty). Again this was with full minors. There has been an increase in the number of trades (which I had suspected based upon the results from the no minor sims) All three sims made 88 trades in 1996. (it was 57-61 with no minors) There has also been a reduction in the frequency in cash being involved but not much. In the 5 sims I did without minors in 2001, cash was involved between 46% and 57% of the deals. In the 3 1996 sims this has dropped to between 38%-43%. The tighter distribution may indicate boundaries of the AI.

This number appears very high. While certainly we are not always privy to the financial parts of deals in the papers, even if the AI believes it's 40%, it seems very high. Also the $3m limit clearly exists. No trade exceeds that amount and it is the most frequent number, even with all the minor leaguers. Again there appears to be no accounts for the club, as the Brewers may get $3m in a deal or the Dodgers/Yankees/etc.

Lets say determine your ability to accept a trade based upon a variety of factors (let's call them value points). The cash has value points (it appears it is the same value points for each team). So I'm trading player A who has 135 (This is a made up number) value points (a combination of factors you consider important about a player [e.g. speed, power, loyalty, years remaining on contract, age, etc] for player B. For the computer to make the deal it wants to be within 10(again made up) points (probably a parameter in the AI). Let's say Player B is 115 value points. With a difference of 20 points the AI won't deal. But wait, cash has value. If we assume $100k per point, if you offer Player B and $1,000,000 (worth 10 vaule points) then you are 10 points apart and the AI makes the deal. To limit this there is an artificial limit of $3M (otherwise the AI would throw all kinds of cash around.).

I believe this is an area the AI needs to be tweaked. While simple in implementation, the reality is that cash is usually NOT why teams trade. Cash is usually somthing the smaller clubs need to make the salary of the acquired player work. I don't know the logic being used here, but I would suspect some iterating logic is needed where different combinations of players are tossed back and forth until both team strategies are met. (example is only break off a negotiation when the value difference is deviating on 4 straight interactions or the value of the trade has not gotten to within 20 points (assuming 10 is the accept range). This made lead to some more creative offers as the trades are pretty static (more on this in later post) in structure. The current use of cash makes for lots of trades, but I question the "realism" of those trades.

One final point, and this may not be a bug but a area in the manual to be clarified. Difficulty appears to have no affect on computer trade. I expect it adjusts the parameter of acceptance on Human v AI trades. So for those trying to cut down trades, you need to reduce frequency as this definitely has an affect, but difficulty does not appear to be impacting computer trades (again more on this later).

Again I would appreciate the thoughts of others here.

[markyah1983]

Currently i have no thoughts but shall get to work on this this evening. Interesting tho.

[Riverdome]

No real thought to add to the thread but it was a very good read. Any time you are able to put into such research is great appreciated.

[Bristolduke]

Unfortunately, work is interferring with my baseball gaming, which happens from time to time. Have have done some simming but mostly I have been trying to analyze the trades. This may be where others can help as well.

In trying to address Raidergoo's view that the sim should be able to make these deals:

Robinson for Pappas, Baldschun and Simpson or
Cincinnati Reds acquire Joe Morgan, Jack Billingham, Denis Menke, Cesar Geronimo and Ed Armbrister from the Houston Astros for Lee May, Tommy Helms, and Jimmy Stewart

I alos this it would be helpful in improving the AI mechanism. I have seen some threads about AI roster moves and signing FAs and there may be a relationship with trading as well, particularly with respect to FAs.

So the set up is 1996 with minor leagues (this will cause some difficulty for others to help evaluate because the minor leaguers will be diferent). My first trade I would like to discuss is Robin Vertura. In all threes 1196 sims he is quickly traded. Why? 1) His listed position is 1B where Frank Thomas plays both players are FA's at the end of the eason; both make about the sam (Thomas $11.8/Venturas 12.6); their loyalty is the same; BUT Ventura is twice as greedy as Thomas. So, the AI GM determines Thomas is a better value to keep. 2) Ventura is also rated at 3B. Tony Phillips is also the 3B. Phillips is not they player of Ventura; Phillips is 36 to Ventura's 28. Phillips make $9M+ has higher loyalty and less greed. Still, the White Sox are rich so money shouldn't be a factor.

It is strange that the sim would rate Phillips over Ventura in this scenario and I am assuming he does only because since Ventura is traded within 20 days of starting the sim, the AI has determined he is expendable. The assumption being if he were traded only once it might be to respond to a great deal, but to be traded in all 3 so fast, the sim must be "trying to move him". So again, the roster/FA logic of the AI will play in the trade logic, so some of the "odd trades may stem from roster management.

Let's look at the deals. Here are the three trades.

1) The Detroit Tigers traded 27-year old left fielder C. Pride to the Chicago White Sox, getting 28-year old first baseman R. Ventura and $3,000,000 in cash in return.

2) The Chicago White Sox traded 28-year old first baseman R. Ventura and $3,000,000 in cash to the San Diego Padres, getting 28-year old reliever A. Ashby and 22-year old minor league third baseman H. Combie in return.

3) The Chicago White Sox traded 28-year old first baseman R. Ventura, 20-year old minor league reliever J. Warner and $3,000,000 in cash to the Detroit Tigers, getting 28-year old catcher J. Flaherty and 21-year old minor league left fielder D. Albright in return.

The fact that all trades send $3M along with Ventura would seem to indicate this is a salary assistance. Of the deals themselves, they don't look great but in the case of #3, the White Sox are weak at C with Krueter and Tolle so Flaherty is an upgrade. The same is true in deal 2 where Ashby is a better starter than their current 3/4/5 starters. (NOTE: I see a lot of starters listed as MRs in the historical pulls - just seems odd). In deal number 1 Pride is not really an upgrade to the White Sox OF nor is it as big of need as other positions.

I'll post other trades up here as well, but I've got to go at the moment. I probably have some real Clunkers, but again the key is where is the AI weak.

Discussions on Ventura trades are encouraged.

[Raidergoo]

Where is Ventura along in his career in regards to arbitration and free agency? It may be that the AI has determined that contract talks are at an impasse and they opted to get something instead of lose him after the season.

[Bristolduke]

Certainly is a factor. He is 28, already has 6 years service, last year of 1 year contract (e.g. his first FA opportunity). His loyalty is 14, so I would have thought he would agreeable to re-sign. The fact that this occurs within the first 20 days of starting the game would indicate to me that not a lot of negotiation has occurred. Trade dates were 3 Jan, 5 Jan, 19 Jan. My guess was the greed factor (14 v Thomas 7) pushed it a bit.

Also, if it was get what you can, why rush in with early trades?

Again, I used this example, in that there may be a correlation with the Roster/FA management, leading to the assumption you have for getting what you can.

These are AI/AI trades, I want to do some testing, time permitting, to see how much I could get for a player.

[John]

I believe i pointed out, in an earlier post, that $3,000,000 is the magic number that the AI almost ALWAYS includes in trades (the post was called, I believe, the $1,290,000 bug - I play on 0.43 Financial settings... at 43% of $3,000,000 the math seems to work). Above are 3 examples, all hitting the same $$$$$. That has to be a bug somewhere.

[TonyJ]

First of all, the game should be rating Ventura over Phillips, that much is obvious. Money shouldn't be a factor, as mentioned, so one can only assume that the AI thinks that Phillips is better. Now, I remember Tony Phillips, he was a good and underrated ballplayer, but Robin Ventura in his prime was a far more valuable commodity than Phillips. So that's obviously a problem.

Secondly, all of those returns are pitiful. Not even not good, they are awful. Pride was a fringe major leaguer, Ashby had a few seasons of usefulness and Flaherty was a no-hit backup catcher. Even assuming that the team felt they had to move him because he was leaving anyway (and, again, they should have rated him over Phillips and kept him), as Bristolduke mentioned they wouldn't have been in such a rush to deal him. More importantly and to the key point, they would have received more value in return even if they felt like they had to move him. A real GM would never settle for such a pitiful return for such a good player. If the offers were that weak they would simply let him go as a FA and take the compensation picks.

I'm no programmer so I have no idea how to fix those things, but it seems like the first problem is somewhat easily resolved (tweaking how the AI values players) while the second may be more difficult.

[JimG]

Originally Posted by Arlie Latham

That is a bad trade.

Jimmy Stewart was well beyond his prime. His MVP seasons of It's a Wonderful Life, Mr. Smith Goes to Washington and The Man Who Shot Liberty Valence are far in the rear view mirror by 1971.

More accurately, Jimmy Stewarts best work was in the Rear Window by this time.

[redmarkYankees]

I believe this possible issue might be significantly affecting trading.

I also think the balance between Overall/Talent/Career may be wrong (or, too impacted by tendencies/game settings/scouting).

Another possibility is that the AI is using Offensive Value & Pitching Value in trade assessment rather than overall value; this would overvalue MRs and CLs, as appears to be the case in a number of odd looking trades. (Pitching value does not differentiate between pitching roles, whereas Overall value does).

[Bristolduke]

A couple of follow ups before my travels this week.

redmarkYankees - I would agree that the issue you reference is a factor. Just for clarification, my tests have all preferences set in the middle, so they should not be influencing anything. (BTW - How do I find the info and against what files can I use SQL?)

I tried to re-sign Ventura and the answer was no, so it is safe to assume the AI knows this and has decided he must be moved to get some value. I also set up a test where I control the White Sox and tried to trade him. Ventura + $3M for Higginson was agreed to instantly, so clearly the AI could have done better than the three trades documented.

It appears the AI's view of teams weaknesses controls the type of player requested and possibly the team sought after. I ran a tests in control of the Wite Sox and no one asked for Ventura. So I can only assume the AI is looking to move him (based upon the analysis I stated above). It may be that the AI scans teams with a weakness in the position he is trading and asks for a player that the AI believes is a weak position for its team. This would be very sound logic. It may also be that this logic is over-riding or not taking into account the value as much as it should. It does appear that the AI could be making stronger offers (based upon the Ventura sample) than it is. While the sample size here is one and insufficient for conclusions on AI logic, I believe that the pattern of player value and weak offers will hold thru the other trades I believe the AI initiates. I'll try to raise those shortly.

There is one more pattern which may be inhibiting the AI's ability to pull off the mega trade and make these weak trades as well. Almost all the deals are one of three types (the exceptions may have more to do with roster problems than being really trade types), these are 1) Major leaguer for Major Leaguer 2) Major Leaguer for Major Leaguer and Minor Leaguer and 3) major Leaguer and Minor Leaguer for Major Leaguer and Minor Leaguer. Obviously we have discussed the inclusion of cash and some of the apparent probelms there. To pull off the trades referenced by Raidergoo, you would have to get out of these structures. So at this point it is still not clear whether the AI would accept either side ot Robinson for Pappas, Baldschun and Simpson, but it does not appear it is NOT capable of offereing either side of this deal.

[redmarkYankees]

There is a utility in the mods forum, I think, by J Altiere which takes the SQL dump and loads into MS Access for you; alternatively, the csv dumps give all the same info.

edit: players.csv gives basic info and name/ID relationship, player value.csv includes all evaluation totals.

I think you're definitely right about team weaknesses; I noticed that when attempting a real solo game, that the AI was much more responsive to offers where I offered players who addressed a team weakness. (If you want to test AI trade responses, try doing it on July 31st or during Winter meetings, when the trade screen offers more immediate clues, though it doesn't suggest players).

On one test I did, I did see a handful of multi-player deals (from memory, 3 or 4 for 1 or 2). That wasn't quite what I was looking at though, so I didn't keep details and I've since deleted the league.